File size: 4,214 Bytes
9425aed | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 | #pragma once
#include <stddef.h>
#include <stdint.h>
#include "rowm_cuda_validation.h"
#ifdef __cplusplus
extern "C" {
#endif
typedef uint64_t CUdeviceptr;
typedef int32_t CUdevice;
typedef int32_t CUdevice_attribute;
typedef int32_t CUresult;
typedef struct CUctx_st* CUcontext;
typedef struct CUmod_st* CUmodule;
typedef struct CUfunc_st* CUfunction;
typedef struct CUstream_st* CUstream;
typedef struct CUevent_st* CUevent;
typedef struct {
uint32_t driver_version;
uint32_t device_ordinal;
uint32_t compute_capability_major;
uint32_t compute_capability_minor;
uint8_t device_uuid[16];
char device_name[64];
uint64_t context_generation;
} sov_cuda_runtime_identity_t;
enum {
CUDA_SUCCESS = 0,
CUDA_ERROR_INVALID_VALUE = 1,
CUDA_ERROR_OUT_OF_MEMORY = 2,
CUDA_ERROR_NOT_INITIALIZED = 3,
CUDA_ERROR_DEINITIALIZED = 4,
CUDA_ERROR_NO_DEVICE = 100,
CUDA_ERROR_INVALID_DEVICE = 101,
CUDA_ERROR_INVALID_IMAGE = 200,
CUDA_ERROR_INVALID_CONTEXT = 201,
CUDA_ERROR_NOT_FOUND = 500,
CUDA_ERROR_UNKNOWN = 999
};
/* Driver and primary-context lifetime */
int sov_cuda_init(void);
void sov_cuda_shutdown(void);
int sov_cuda_is_initialized(void);
uint64_t sov_cuda_context_generation(void);
CUresult sov_cuda_get_runtime_identity(sov_cuda_runtime_identity_t* identity_out);
/* Checked access to the dynamically resolved CUDA Driver API */
CUresult sov_cuda_module_load_data(CUmodule* module, const void* image);
CUresult sov_cuda_module_get_function(CUfunction* function, CUmodule module,
const char* name);
CUresult sov_cuda_module_get_global(CUdeviceptr* device_ptr, size_t* bytes,
CUmodule module, const char* name);
CUresult sov_cuda_module_unload(CUmodule module);
/*
* The raw launch boundary is fail-closed: even callers that bypass the
* higher-level dispatchers need a ROWM-NR commit for the active context.
*/
CUresult sov_cuda_launch_kernel(CUfunction function,
unsigned int grid_x,
unsigned int grid_y,
unsigned int grid_z,
unsigned int block_x,
unsigned int block_y,
unsigned int block_z,
unsigned int shared_mem_bytes,
CUstream stream,
void** kernel_params,
void** extra);
CUresult sov_cuda_mem_alloc(CUdeviceptr* device_ptr, size_t bytes);
CUresult sov_cuda_mem_free(CUdeviceptr device_ptr);
int sov_cuda_memcpy_h2d(void* device_dst, const void* host_src, size_t bytes);
CUresult sov_cuda_memcpy_d2h(void* host_dst, CUdeviceptr device_src, size_t bytes);
/* Power state device symbol binding */
CUresult sov_cuda_register_power_state_device(CUdeviceptr device_ptr);
CUresult sov_cuda_sync_power_state(void);
/* PTX module dispatch */
int sov_cuda_kernels_init(void);
int sov_cuda_kernels_authorize_rowm(sov_rowm_commit_cuda_validation_fn commit_fn,
void* rowm_context);
void sov_cuda_kernels_shutdown(void);
int sov_cuda_rmsnorm_fused(CUdeviceptr x, CUdeviceptr weight,
CUdeviceptr out, uint32_t element_count);
int sov_cuda_silu_fused(CUdeviceptr x, CUdeviceptr out, uint32_t element_count);
/* GEMM dispatch */
int sov_cuda_gemm_init(void);
void sov_cuda_gemm_shutdown(void);
int sov_cuda_gemm(CUdeviceptr A, CUdeviceptr B, CUdeviceptr C,
int M, int N, int K);
int sov_cuda_gemm_ex(CUdeviceptr A, CUdeviceptr B, CUdeviceptr C,
int M, int N, int K, int lda, int ldb, int ldc);
/* Compatibility entry points retained for the existing RTX public API */
int sov_cuda_load_ptx(const char* ptx_data, unsigned int ptx_size,
void** module_out);
void* sov_cuda_malloc(size_t bytes);
#ifdef __cplusplus
}
#endif
|