/
usr
/
local
/
lib64
/
python3.6
/
site-packages
/
torch
/
include
/
THC
/
/usr/local/lib64/python3.6/site-packages/torch/include/THC
mkdir
upload
Name
Size
Mode
Actions
generic/
-
0755
rm
THC.h
328
0644
edit
dl
rm
THCAllocator.h
370
0644
edit
dl
rm
THCAsmUtils.cuh
3444
0644
edit
dl
rm
THCAtomics.cuh
13094
0644
edit
dl
rm
THCCachingHostAllocator.h
1254
0644
edit
dl
rm
THCDeviceTensor-inl.cuh
11515
0644
edit
dl
rm
THCDeviceTensor.cuh
16160
0644
edit
dl
rm
THCDeviceTensorUtils-inl.cuh
4497
0644
edit
dl
rm
THCDeviceTensorUtils.cuh
2720
0644
edit
dl
rm
THCDeviceUtils.cuh
942
0644
edit
dl
rm
THCGeneral.h
2741
0644
edit
dl
rm
THCGeneral.hpp
743
0644
edit
dl
rm
THCGenerateAllTypes.h
958
0644
edit
dl
rm
THCGenerateBFloat16Type.h
515
0644
edit
dl
rm
THCGenerateBoolType.h
449
0644
edit
dl
rm
THCGenerateByteType.h
419
0644
edit
dl
rm
THCGenerateCharType.h
418
0644
edit
dl
rm
THCGenerateComplexDoubleType.h
533
0644
edit
dl
rm
THCGenerateComplexFloatType.h
526
0644
edit
dl
rm
THCGenerateComplexTypes.h
298
0644
edit
dl
rm
THCGenerateDoubleType.h
464
0644
edit
dl
rm
THCGenerateFloatType.h
550
0644
edit
dl
rm
THCGenerateFloatTypes.h
779
0644
edit
dl
rm
THCGenerateHalfType.h
481
0644
edit
dl
rm
THCGenerateIntType.h
414
0644
edit
dl
rm
THCGenerateLongType.h
419
0644
edit
dl
rm
THCGenerateShortType.h
424
0644
edit
dl
rm
THCIntegerDivider.cuh
4095
0644
edit
dl
rm
THCNumerics.cuh
19813
0644
edit
dl
rm
THCScanUtils.cuh
4789
0644
edit
dl
rm
THCSleep.h
232
0644
edit
dl
rm
THCStorage.h
487
0644
edit
dl
rm
THCStorage.hpp
846
0644
edit
dl
rm
THCStorageCopy.h
466
0644
edit
dl
rm
THCTensor.h
623
0644
edit
dl
rm
THCTensor.hpp
1070
0644
edit
dl
rm
THCTensorCopy.h
467
0644
edit
dl
rm
THCTensorCopy.hpp
600
0644
edit
dl
rm
THCTensorMathReduce.cuh
664
0644
edit
dl
rm
THCThrustAllocator.cuh
618
0644
edit
dl
rm
Edit:
/usr/local/lib64/python3.6/site-packages/torch/include/THC/THCAsmUtils.cuh
(3444B)
#ifndef THC_ASM_UTILS_INC #define THC_ASM_UTILS_INC // Collection of direct PTX functions template <typename T> struct Bitfield {}; template <> struct Bitfield<unsigned int> { static __device__ __forceinline__ unsigned int getBitfield(unsigned int val, int pos, int len) { #if defined(__HIP_PLATFORM_HCC__) pos &= 0xff; len &= 0xff; unsigned int m = (1u << len) - 1u; return (val >> pos) & m; #else unsigned int ret; asm("bfe.u32 %0, %1, %2, %3;" : "=r"(ret) : "r"(val), "r"(pos), "r"(len)); return ret; #endif } static __device__ __forceinline__ unsigned int setBitfield(unsigned int val, unsigned int toInsert, int pos, int len) { #if defined(__HIP_PLATFORM_HCC__) pos &= 0xff; len &= 0xff; unsigned int m = (1u << len) - 1u; toInsert &= m; toInsert <<= pos; m <<= pos; return (val & ~m) | toInsert; #else unsigned int ret; asm("bfi.b32 %0, %1, %2, %3, %4;" : "=r"(ret) : "r"(toInsert), "r"(val), "r"(pos), "r"(len)); return ret; #endif } }; template <> struct Bitfield<uint64_t> { static __device__ __forceinline__ uint64_t getBitfield(uint64_t val, int pos, int len) { #if defined(__HIP_PLATFORM_HCC__) pos &= 0xff; len &= 0xff; uint64_t m = (1u << len) - 1u; return (val >> pos) & m; #else uint64_t ret; asm("bfe.u64 %0, %1, %2, %3;" : "=l"(ret) : "l"(val), "r"(pos), "r"(len)); return ret; #endif } static __device__ __forceinline__ uint64_t setBitfield(uint64_t val, uint64_t toInsert, int pos, int len) { #if defined(__HIP_PLATFORM_HCC__) pos &= 0xff; len &= 0xff; uint64_t m = (1u << len) - 1u; toInsert &= m; toInsert <<= pos; m <<= pos; return (val & ~m) | toInsert; #else uint64_t ret; asm("bfi.b64 %0, %1, %2, %3, %4;" : "=l"(ret) : "l"(toInsert), "l"(val), "r"(pos), "r"(len)); return ret; #endif } }; __device__ __forceinline__ int getLaneId() { #if defined(__HIP_PLATFORM_HCC__) return __lane_id(); #else int laneId; asm("mov.s32 %0, %%laneid;" : "=r"(laneId) ); return laneId; #endif } #if defined(__HIP_PLATFORM_HCC__) __device__ __forceinline__ unsigned long long int getLaneMaskLt() { const std::uint64_t m = (1ull << getLaneId()) - 1ull; return m; } #else __device__ __forceinline__ unsigned getLaneMaskLt() { unsigned mask; asm("mov.u32 %0, %%lanemask_lt;" : "=r"(mask)); return mask; } #endif #if defined (__HIP_PLATFORM_HCC__) __device__ __forceinline__ unsigned long long int getLaneMaskLe() { std::uint64_t m = UINT64_MAX >> (sizeof(std::uint64_t) * CHAR_BIT - (getLaneId() + 1)); return m; } #else __device__ __forceinline__ unsigned getLaneMaskLe() { unsigned mask; asm("mov.u32 %0, %%lanemask_le;" : "=r"(mask)); return mask; } #endif #if defined(__HIP_PLATFORM_HCC__) __device__ __forceinline__ unsigned long long int getLaneMaskGt() { const std::uint64_t m = getLaneMaskLe(); return m ? ~m : m; } #else __device__ __forceinline__ unsigned getLaneMaskGt() { unsigned mask; asm("mov.u32 %0, %%lanemask_gt;" : "=r"(mask)); return mask; } #endif #if defined(__HIP_PLATFORM_HCC__) __device__ __forceinline__ unsigned long long int getLaneMaskGe() { const std::uint64_t m = getLaneMaskLt(); return ~m; } #else __device__ __forceinline__ unsigned getLaneMaskGe() { unsigned mask; asm("mov.u32 %0, %%lanemask_ge;" : "=r"(mask)); return mask; } #endif #endif // THC_ASM_UTILS_INC
Save
cmd:
run