38 #include <hip/hip_fp16.h>
40 #if __has_include(<half/half.hpp>)
41 #include <half/half.hpp>
45 using halfhpp = half_float::half;
46 typedef halfhpp Rpp16f;
50 #elif defined(__x86_64__) || defined(_M_X64)
51 #include <immintrin.h>
52 #include <smmintrin.h>
53 #include <x86intrin.h>
57 #if defined(__HIP_DEVICE_COMPILE__)
58 #define SIMDE_NO_NATIVE
60 #define SIMDE_ENABLE_NATIVE_ALIASES
61 #include <simde/x86/avx2.h>
62 #include <simde/x86/f16c.h>
63 #include <simde/x86/fma.h>
65 #ifndef _MM_FROUND_NO_EXC
66 #define _MM_FROUND_NO_EXC 0x08
72 #define RPP_MIN_8U (0)
74 #define RPP_MAX_8U (255)
76 #define RPPT_MAX_DIMS (5)
79 #define RPPT_MAX_DIMS_SAMPLE (4)
81 #define RPPT_MAX_AUDIO_CHANNELS (16)
85 #define RPP_RETURN_IF_ERROR(expr) \
87 RppStatus _rpp_status = (expr); \
88 if (_rpp_status != RPP_SUCCESS) return _rpp_status; \
92 #include <hip/hip_runtime.h>
93 #define RPP_HOST_DEVICE __host__ __device__
97 #define RPP_HIP_RETURN_IF_ERROR(expr) \
99 hipError_t _rpp_hip_err = (expr); \
100 if (_rpp_hip_err != hipSuccess) { \
101 fprintf(stderr, "HIP runtime error: %s returned %d at %s:%d", #expr, _rpp_hip_err, \
102 __FILE__, __LINE__); \
103 return RPP_ERROR_HIP_RUNTIME; \
109 #define HIP_CHECK_LAUNCH_RETURN() \
111 hipError_t status = hipGetLastError(); \
112 if (status != hipSuccess) { \
113 fprintf(stderr, "HIP kernel launch error: returned %d at %s:%d", status, __FILE__, \
115 return RPP_ERROR_HIP_LAUNCH; \
119 #define RPP_HOST_DEVICE
122 const float ONE_OVER_6 = 1.0f / 6;
123 const float ONE_OVER_3 = 1.0f / 3;
124 const float TWO_OVER_3 = 2.0f / 3;
125 const float ONE_OVER_255 = 1.0f / 255;
126 const uint MMS_MAX_SCRATCH_MEMORY = 115293120;
128 const uint SPECTROGRAM_MAX_SCRATCH_MEMORY =
134 #define DROPOUT_FIXED_SEED 42
135 #define RANDOM_ERASE_NOISE_BUFFER_SIDE \
267 return "RPP_SUCCESS";
271 return "RPP_ERROR_INVALID_ARGUMENTS";
273 return "RPP_ERROR_LOW_OFFSET";
275 return "RPP_ERROR_ZERO_DIVISION";
277 return "RPP_ERROR_HIGH_SRC_DIMENSION";
279 return "RPP_ERROR_NOT_IMPLEMENTED";
281 return "RPP_ERROR_INVALID_SRC_CHANNELS";
283 return "RPP_ERROR_INVALID_DST_CHANNELS";
285 return "RPP_ERROR_INVALID_SRC_LAYOUT";
287 return "RPP_ERROR_INVALID_DST_LAYOUT";
289 return "RPP_ERROR_INVALID_SRC_DATATYPE";
291 return "RPP_ERROR_INVALID_DST_DATATYPE";
293 return "RPP_ERROR_INVALID_SRC_OR_DST_DATATYPE";
295 return "RPP_ERROR_INSUFFICIENT_DST_BUFFER_LENGTH";
297 return "RPP_ERROR_INVALID_PARAMETER_DATATYPE";
299 return "RPP_ERROR_NOT_ENOUGH_MEMORY";
301 return "RPP_ERROR_OUT_OF_BOUND_SRC_ROI";
303 return "RPP_ERROR_LAYOUT_MISMATCH";
305 return "RPP_ERROR_INVALID_CHANNELS";
307 return "RPP_ERROR_INVALID_OUTPUT_TILE_LENGTH";
309 return "RPP_ERROR_OUT_OF_BOUND_SHARED_MEMORY_SIZE";
311 return "RPP_ERROR_OUT_OF_BOUND_SCRATCH_MEMORY_SIZE";
313 return "RPP_ERROR_INVALID_SRC_DIMS";
315 return "RPP_ERROR_INVALID_DST_DIMS";
317 return "RPP_ERROR_INVALID_DIM_LENGTHS";
319 return "RPP_ERROR_INVALID_AXIS";
321 return "RPP_ERROR_INCOMPATIBLE_BACKEND";
323 return "RPP_ERROR_HIP_LAUNCH";
325 return "RPP_ERROR_HIP_RUNTIME";
328 return "RPP_STATUS_UNKNOWN";
334 typedef enum { RPP_HOST_BACKEND, RPP_HIP_BACKEND }
RppBackend;
340 rppStatusSuccess = 0,
341 rppStatusBadParm = -1,
342 rppStatusUnknownError = -2,
343 rppStatusNotInitialized = -3,
344 rppStatusInvalidValue = -4,
345 rppStatusAllocFailed = -5,
346 rppStatusInternalError = -6,
347 rppStatusNotImplemented = -7,
348 rppStatusUnsupportedOp = -8,
356 RPP_TENSOR_OP_SUBTRACT,
357 RPP_TENSOR_OP_MULTIPLY,
364 typedef enum { RPP_TENSOR_OP_AND, RPP_TENSOR_OP_OR, RPP_TENSOR_OP_XOR }
RpptBitwiseOp;
374 RPP_BROADCAST_ENABLE,
375 RPP_BROADCAST_DISABLE
497 NEAREST_NEIGHBOR = 0,
580 int roiWidth, roiHeight;
589 int roiWidth, roiHeight, roiDepth;
635 : dataType(dataType), n(n), c(c), h(h), w(w), layout(layout) {
636 if (layout == NCHW) {
637 strides.nStride = c * h * w;
638 strides.cStride = h * w;
641 }
else if (layout == NHWC) {
642 strides.nStride = h * w * c;
643 strides.hStride = w * c;
651 : offsetInBytes(offsetInBytes),
731 float boxMullerExtra;
763 switch (interpolationType) {
764 case RpptInterpolationType::BICUBIC: {
768 case RpptInterpolationType::LANCZOS: {
769 if (in_size > out_size) {
770 this->radius = 3.0f * scaleRatio;
771 this->scale = (1 / scaleRatio);
776 case RpptInterpolationType::GAUSSIAN: {
777 if (in_size > out_size) {
778 this->radius = scaleRatio;
779 this->scale = (1 / scaleRatio);
783 case RpptInterpolationType::TRIANGULAR: {
784 if (in_size > out_size) {
785 this->radius = scaleRatio;
786 this->scale = (1 / scaleRatio);
796 this->size = std::ceil(2 * this->radius);
811 Rpp32f locRaw = x * scale + center;
812 Rpp32s locFloor = std::floor(locRaw);
813 Rpp32f weight = locRaw - locFloor;
814 locFloor = std::max(std::min(locFloor, lookupSize - 2), 0);
815 Rpp32f current = lookup[locFloor];
816 Rpp32f next = lookup[locFloor + 1];
817 return current + weight * (next - current);
820 inline __m128 operator()(__m128 x) {
821 __m128 pLocRaw = _mm_add_ps(_mm_mul_ps(x, pScale), pCenter);
822 __m128i pxLocFloor = _mm_cvttps_epi32(pLocRaw);
823 __m128 pLocFloor = _mm_cvtepi32_ps(pxLocFloor);
824 __m128 pWeight = _mm_sub_ps(pLocRaw, pLocFloor);
826 _mm_storeu_si128(
reinterpret_cast<__m128i*
>(idx), pxLocFloor);
828 _mm_setr_ps(lookup[idx[0]], lookup[idx[1]], lookup[idx[2]], lookup[idx[3]]);
829 __m128 pNext = _mm_setr_ps(lookup[idx[0] + 1], lookup[idx[1] + 1], lookup[idx[2] + 1],
831 return _mm_add_ps(pCurrent, _mm_mul_ps(pWeight, _mm_sub_ps(pNext, pCurrent)));
834 Rpp32f scale = 1, center = 1;
835 Rpp32s lobes = 0, coeffs = 0;
837 Rpp32f* lookupPinned =
nullptr;
838 std::vector<Rpp32f> lookup;
839 __m128 pCenter, pScale;
847 inline RPP_HOST_DEVICE
virtual Rpp32f hz_to_mel(
Rpp32f hz) = 0;
848 inline RPP_HOST_DEVICE
virtual Rpp32f mel_to_hz(
Rpp32f mel) = 0;
857 return 1127.0f * std::log(1.0f + (hz / 700.0f));
860 return 700.0f * (std::exp(mel / 1127.0f) - 1.0f);
872 const Rpp32f fsp = 66.666667f;
873 const Rpp32f minLogHz = 1000.0;
874 const Rpp32f minLogMel = (minLogHz - freqLow) / fsp;
875 const Rpp32f stepLog = 0.068751777;
877 const Rpp32f invMinLogHz = 0.001f;
878 const Rpp32f invStepLog = 1.0f / stepLog;
879 const Rpp32f invFsp = 1.0f / fsp;
884 mel = minLogMel + std::log(hz * invMinLogHz) * invStepLog;
886 mel = (hz - freqLow) * invFsp;
893 if (mel >= minLogMel)
894 hz = minLogHz * std::exp(stepLog * (mel - minLogMel));
896 hz = freqLow + mel * fsp;
910 Rpp32f* scratchBufferHost;
unsigned int Rpp32u
32 bit unsigned int
Definition: rppdefs.h:150
int Rpp32s
32 bit signed int
Definition: rppdefs.h:152
signed char Rpp8s
8 bit signed char
Definition: rppdefs.h:144
RpptMelScaleFormula
RPPT Mel Scale Formula.
Definition: rppdefs.h:527
unsigned short Rpp16u
16 bit unsigned short
Definition: rppdefs.h:146
unsigned char Rpp8u
8 bit unsigned char
Definition: rppdefs.h:142
void * RppPtr_t
void pointer
Definition: rppdefs.h:162
RpptLayout
RPPT Tensor layout type enum.
Definition: rppdefs.h:425
RpptBayerPattern
Bayer colour filter array pattern for rppt_bayer_to_rgb.
Definition: rppdefs.h:465
RpptRoiType
RPPT Tensor 2D ROI type enum.
Definition: rppdefs.h:475
RpptDataType
RPPT Tensor datatype enum.
Definition: rppdefs.h:420
RpptImageBorderType
RPPT Image Border Type.
Definition: rppdefs.h:518
struct RpptResamplingWindow RpptResamplingWindow
RPPT Tensor RpptResamplingWindow type struct.
double Rpp64f
64 bit double
Definition: rppdefs.h:160
RpptImageBorderEdge
RPPT Image Border Edge type enum.
Definition: rppdefs.h:513
RpptAudioBorderType
RPPT Audio Border Type.
Definition: rppdefs.h:508
size_t RppSize_t
size_t
Definition: rppdefs.h:164
RpptBitwiseOp
RPPT Broadcast Mode type enum.
Definition: rppdefs.h:364
RpptColorStandard
Color space standard for NV12 to RGB (YUV/YCbCr to RGB matrix).
Definition: rppdefs.h:439
struct GenericFilter GenericFilter
RPPT Tensor GenericFilter type struct.
RppStatus
RPP RppStatus type enums.
Definition: rppdefs.h:169
RpptColorRange
Color range (luma/chroma legal levels) for NV12 to RGB.
Definition: rppdefs.h:455
RpptRoi3DType
RPPT Tensor 3D ROI type enum.
Definition: rppdefs.h:483
const char * rppGetStatusString(RppStatus status) noexcept
Get the string representation of a RPP status code.
Definition: rppdefs.h:262
unsigned long long Rpp64u
64 bit unsigned long long
Definition: rppdefs.h:154
#define RPPT_MAX_DIMS
RPP maximum dimensions in tensor .
Definition: rppdefs.h:76
RpptInterpolationType
RPPT Tensor interpolation type enum.
Definition: rppdefs.h:496
RpptSubpixelLayout
RPPT Tensor subpixel layout type enum.
Definition: rppdefs.h:491
RpptOp
RPP Tensor Operations type enum.
Definition: rppdefs.h:354
short Rpp16s
16 bit signed short
Definition: rppdefs.h:148
float Rpp32f
32 bit float
Definition: rppdefs.h:158
RppBackend
RPP RppBackend type enums.
Definition: rppdefs.h:334
long long Rpp64s
64 bit long long
Definition: rppdefs.h:156
rppStatus_t
RPP rppStatus_t type enums.
Definition: rppdefs.h:339
@ RpptBayerPattern_RGGB
Definition: rppdefs.h:466
@ RpptBayerPattern_BGGR
Definition: rppdefs.h:467
@ RpptBayerPattern_GBRG
Definition: rppdefs.h:469
@ RpptBayerPattern_GRBG
Definition: rppdefs.h:468
@ RpptColorStandard_FCC
Definition: rppdefs.h:441
@ RpptColorStandard_SMPTE240M
Definition: rppdefs.h:444
@ RpptColorStandard_BT601
Definition: rppdefs.h:443
@ RpptColorStandard_BT470BG
Definition: rppdefs.h:442
@ RpptColorStandard_BT709
Definition: rppdefs.h:440
@ RpptColorStandard_BT2020_NCL
Definition: rppdefs.h:445
@ RpptColorStandard_BT2020_CL
Definition: rppdefs.h:446
@ RPP_ERROR_NOT_IMPLEMENTED
Function variant requested is not implemented / unsupported.
Definition: rppdefs.h:186
@ RPP_ERROR_INVALID_DST_DATATYPE
Invalid dst tensor datatype. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:204
@ RPP_ERROR_INVALID_OUTPUT_TILE_LENGTH
Invalid output tile length (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:225
@ RPP_ERROR_INVALID_DIM_LENGTHS
Length in some dimensions are invalid. (Needs to adhere to function specification....
Definition: rppdefs.h:240
@ RPP_ERROR_OUT_OF_BOUND_SRC_ROI
Out of bound source ROI.
Definition: rppdefs.h:217
@ 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:228
@ RPP_ERROR_INCOMPATIBLE_BACKEND
The user specified backend is not compatible with the initialized handle group_rppdefs.
Definition: rppdefs.h:246
@ RPP_ERROR_HIP_LAUNCH
HIP/GPU runtime or kernel launch error.
Definition: rppdefs.h:248
@ RPP_ERROR_ZERO_DIVISION
Arguments provided will result in zero division error.
Definition: rppdefs.h:180
@ RPP_ERROR_INVALID_ARGUMENTS
One or more arguments invalid. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:176
@ RPP_ERROR_LAYOUT_MISMATCH
src and dst layout mismatch
Definition: rppdefs.h:219
@ RPP_ERROR_INVALID_CHANNELS
Number of channels is invalid. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:222
@ RPP_ERROR_INVALID_SRC_CHANNELS
Invalid src tensor number of channels. (Needs to adhere to function specification....
Definition: rppdefs.h:189
@ RPP_ERROR_INVALID_SRC_OR_DST_DATATYPE
Invalid src/dst tensor datatype. (Needs to adhere to function specification.)
Definition: rppdefs.h:207
@ RPP_ERROR_INVALID_PARAMETER_DATATYPE
Invalid datatype.
Definition: rppdefs.h:212
@ RPP_ERROR_NOT_ENOUGH_MEMORY
Not enough memory to write outputs, as per dim-lengths and strides set in descriptor.
Definition: rppdefs.h:215
@ RPP_ERROR
Unspecified error.
Definition: rppdefs.h:173
@ RPP_ERROR_INVALID_SRC_DIMS
Number of src dims is invalid. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:234
@ 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:231
@ RPP_ERROR_INVALID_DST_LAYOUT
Invalid dst tensor layout. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:198
@ RPP_ERROR_INVALID_AXIS
The specified axis is invalid or out of range. (Needs to adhere to function specification....
Definition: rppdefs.h:243
@ RPP_ERROR_INVALID_DST_CHANNELS
Invalid dst tensor number of channels. (Needs to adhere to function specification....
Definition: rppdefs.h:192
@ RPP_ERROR_HIP_RUNTIME
Internal HIP/GPU runtime error.
Definition: rppdefs.h:250
@ RPP_ERROR_INVALID_SRC_DATATYPE
Invalid src tensor datatype. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:201
@ RPP_ERROR_INVALID_SRC_LAYOUT
Invalid src tensor layout. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:195
@ RPP_ERROR_HIGH_SRC_DIMENSION
Src tensor / src ROI dimension too high. (Needs to adhere to function specification....
Definition: rppdefs.h:183
@ RPP_SUCCESS
No error.
Definition: rppdefs.h:171
@ RPP_ERROR_INSUFFICIENT_DST_BUFFER_LENGTH
Insufficient dst buffer length provided. (Needs to adhere to function specification....
Definition: rppdefs.h:210
@ RPP_ERROR_INVALID_DST_DIMS
Number of dst dims is invalid. (Needs to adhere to function specification.) group_rppdefs.
Definition: rppdefs.h:237
@ RPP_ERROR_LOW_OFFSET
Low tensor offsetInBytes provided for src/dst tensor.
Definition: rppdefs.h:178
@ RpptColorRange_FULL
Definition: rppdefs.h:457
@ RpptColorRange_STUDIO
Definition: rppdefs.h:456
RpptBroadcastMode
Broadcast support mode for RPP Tensor operations.
Definition: rppdefs.h:373
Base class for Mel scale conversions.
Definition: rppdefs.h:845
RPPT Tensor GenericFilter type struct.
Definition: rppdefs.h:757
Derived class for HTK Mel scale conversions.
Definition: rppdefs.h:855
RPP initialize handle.
Definition: rppdefs.h:954
RPP 24 float vector.
Definition: rppdefs.h:411
RPP 6 float vector.
Definition: rppdefs.h:390
RPP 9 float vector.
Definition: rppdefs.h:397
RPP 24 signed int vector.
Definition: rppdefs.h:404
RPP layout params.
Definition: rppdefs.h:382
RPPT Tensor 2D bilinear neighborhood 32-bit signed int 8-length-vectors type struct.
Definition: rppdefs.h:737
RPPT Tensor 2D bilinear neighborhood 32-bit float 8-length-vectors type struct.
Definition: rppdefs.h:747
RPPT Tensor Channel Offsets struct.
Definition: rppdefs.h:561
RPPT Tensor descriptor type struct.
Definition: rppdefs.h:624
RPPT Tensor 32-bit float RGB type struct.
Definition: rppdefs.h:686
RPPT Tensor 2D 32-bit float vector type struct.
Definition: rppdefs.h:703
RPPT Tensor Generic descriptor type struct.
Definition: rppdefs.h:665
RPPT Tensor 2D image patch dimensions type struct.
Definition: rppdefs.h:711
RPPT Image 2D cartesian point type struct.
Definition: rppdefs.h:536
RPPT Image 3D point type struct.
Definition: rppdefs.h:544
RPPT Tensor 8-bit uchar RGB type struct.
Definition: rppdefs.h:677
RPPT Tensor RpptResamplingWindow type struct.
Definition: rppdefs.h:803
RPPT Tensor 3D ROI LTFRBB struct.
Definition: rppdefs.h:570
RPPT Tensor 2D ROI LTRB struct.
Definition: rppdefs.h:553
RPPT Tensor 2D ROI XYWH struct.
Definition: rppdefs.h:578
RPPT Tensor 3D ROI XYZWHD struct.
Definition: rppdefs.h:587
RPPT Tensor strides type struct.
Definition: rppdefs.h:614
RPPT Tensor 2D 32-bit uint vector type struct.
Definition: rppdefs.h:695
RPPT Tensor random number generator state (xorwow box muller state) type struct.
Definition: rppdefs.h:727
RPPT Tensor random number generator state (xorwow state) type struct.
Definition: rppdefs.h:719
Derived class for Slaney Mel scale conversions.
Definition: rppdefs.h:870
RPP HIP 32-bit float memory.
Definition: rppdefs.h:920
RPP HOST memory type struct.
Definition: rppdefs.h:909
RPP HIP memory management type struct.
Definition: rppdefs.h:927
RPP HIP-HOST memory management.
Definition: rppdefs.h:935
RPPT Tensor 3D ROI union.
Definition: rppdefs.h:605
RPPT Tensor 2D ROI union.
Definition: rppdefs.h:596