|
| 1 | + |
| 2 | +#ifndef HEADER_ACLRTLAUNCH_ROPE_CUSTOM_FALSE_BFLOAT16_T_HKERNEL_H_ |
| 3 | +#define HEADER_ACLRTLAUNCH_ROPE_CUSTOM_FALSE_BFLOAT16_T_HKERNEL_H_ |
| 4 | + |
| 5 | + |
| 6 | + |
| 7 | +extern "C" uint32_t aclrtlaunch_rope_custom_false_bfloat16_t(uint32_t blockDim, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum); |
| 8 | + |
| 9 | +inline uint32_t rope_custom_false_bfloat16_t(uint32_t blockDim, void* hold, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum) |
| 10 | +{ |
| 11 | + (void)hold; |
| 12 | + return aclrtlaunch_rope_custom_false_bfloat16_t(blockDim, stream, positions, queryDst, keyDst, query, key, cosSinCache, rotDim, queryStride, keyStride, dstQueryStride, dstKeyStride, numHeads, numKvHeads, headSize, numTokens, loopNum, coreNum); |
| 13 | +} |
| 14 | + |
| 15 | +#endif |
| 16 | + |
| 17 | +#ifndef HEADER_ACLRTLAUNCH_ROPE_CUSTOM_FALSE_HALF_HKERNEL_H_ |
| 18 | +#define HEADER_ACLRTLAUNCH_ROPE_CUSTOM_FALSE_HALF_HKERNEL_H_ |
| 19 | + |
| 20 | + |
| 21 | + |
| 22 | +extern "C" uint32_t aclrtlaunch_rope_custom_false_half(uint32_t blockDim, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum); |
| 23 | + |
| 24 | +inline uint32_t rope_custom_false_half(uint32_t blockDim, void* hold, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum) |
| 25 | +{ |
| 26 | + (void)hold; |
| 27 | + return aclrtlaunch_rope_custom_false_half(blockDim, stream, positions, queryDst, keyDst, query, key, cosSinCache, rotDim, queryStride, keyStride, dstQueryStride, dstKeyStride, numHeads, numKvHeads, headSize, numTokens, loopNum, coreNum); |
| 28 | +} |
| 29 | + |
| 30 | +#endif |
| 31 | + |
| 32 | +#ifndef HEADER_ACLRTLAUNCH_ROPE_CUSTOM_TRUE_BFLOAT16_T_HKERNEL_H_ |
| 33 | +#define HEADER_ACLRTLAUNCH_ROPE_CUSTOM_TRUE_BFLOAT16_T_HKERNEL_H_ |
| 34 | + |
| 35 | + |
| 36 | + |
| 37 | +extern "C" uint32_t aclrtlaunch_rope_custom_true_bfloat16_t(uint32_t blockDim, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum); |
| 38 | + |
| 39 | +inline uint32_t rope_custom_true_bfloat16_t(uint32_t blockDim, void* hold, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum) |
| 40 | +{ |
| 41 | + (void)hold; |
| 42 | + return aclrtlaunch_rope_custom_true_bfloat16_t(blockDim, stream, positions, queryDst, keyDst, query, key, cosSinCache, rotDim, queryStride, keyStride, dstQueryStride, dstKeyStride, numHeads, numKvHeads, headSize, numTokens, loopNum, coreNum); |
| 43 | +} |
| 44 | + |
| 45 | +#endif |
| 46 | + |
| 47 | +#ifndef HEADER_ACLRTLAUNCH_ROPE_CUSTOM_TRUE_HALF_HKERNEL_H_ |
| 48 | +#define HEADER_ACLRTLAUNCH_ROPE_CUSTOM_TRUE_HALF_HKERNEL_H_ |
| 49 | + |
| 50 | + |
| 51 | + |
| 52 | +extern "C" uint32_t aclrtlaunch_rope_custom_true_half(uint32_t blockDim, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum); |
| 53 | + |
| 54 | +inline uint32_t rope_custom_true_half(uint32_t blockDim, void* hold, void* stream, void* positions, void* queryDst, void* keyDst, void* query, void* key, void* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, const int headSize, const int64_t numTokens, const int loopNum, const int coreNum) |
| 55 | +{ |
| 56 | + (void)hold; |
| 57 | + return aclrtlaunch_rope_custom_true_half(blockDim, stream, positions, queryDst, keyDst, query, key, cosSinCache, rotDim, queryStride, keyStride, dstQueryStride, dstKeyStride, numHeads, numKvHeads, headSize, numTokens, loopNum, coreNum); |
| 58 | +} |
| 59 | + |
| 60 | +#endif |
0 commit comments