issue/1033 - replace __C with __INFINI_C

src/infiniop/ops/logsoftmax/operator.cc

@@ -15,7 +15,7 @@
 // #include "ascend/logsoftmax_ascend.h"
 #endif
 
-__C infiniStatus_t infiniopCreateLogSoftmaxDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateLogSoftmaxDescriptor(
     infiniopHandle_t handle,
     infiniopLogSoftmaxDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -56,7 +56,7 @@ __C infiniStatus_t infiniopCreateLogSoftmaxDescriptor(
     }
 }
 
-__C infiniStatus_t infiniopGetLogSoftmaxWorkspaceSize(infiniopLogSoftmaxDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetLogSoftmaxWorkspaceSize(infiniopLogSoftmaxDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                      \
     case CASE:                                                                                    \
@@ -90,7 +90,7 @@ __C infiniStatus_t infiniopGetLogSoftmaxWorkspaceSize(infiniopLogSoftmaxDescript
     }
 }
 
-__C infiniStatus_t infiniopLogSoftmax(
+__INFINI_C infiniStatus_t infiniopLogSoftmax(
     infiniopLogSoftmaxDescriptor_t desc,
     void *workspace, size_t workspace_size,
     void *y,
@@ -129,7 +129,7 @@ __C infiniStatus_t infiniopLogSoftmax(
     }
 }
 
-__C infiniStatus_t infiniopDestroyLogSoftmaxDescriptor(infiniopLogSoftmaxDescriptor_t desc) {
+__INFINI_C infiniStatus_t infiniopDestroyLogSoftmaxDescriptor(infiniopLogSoftmaxDescriptor_t desc) {
 
 #define DESTROY(CASE, NAMESPACE)                                                \
     case CASE:                                                                  \

src/infiniop/ops/lp_norm/operator.cc

@@ -6,7 +6,7 @@
 #include "nvidia/lp_norm_nvidia.cuh"
 #endif
 
-__C infiniStatus_t infiniopCreateLPNormDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateLPNormDescriptor(
     infiniopHandle_t handle,
     infiniopLPNormDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t output_desc,
@@ -47,7 +47,7 @@ __C infiniStatus_t infiniopCreateLPNormDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetLPNormWorkspaceSize(infiniopLPNormDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetLPNormWorkspaceSize(infiniopLPNormDescriptor_t desc, size_t *size) {
 #define GET(CASE, NAMESPACE)                                                                   \
     case CASE:                                                                                 \
         *size = reinterpret_cast<op::lp_norm::NAMESPACE::Descriptor *>(desc)->workspaceSize(); \
@@ -75,7 +75,7 @@ __C infiniStatus_t infiniopGetLPNormWorkspaceSize(infiniopLPNormDescriptor_t des
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopLPNorm(
+__INFINI_C infiniStatus_t infiniopLPNorm(
     infiniopLPNormDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -114,7 +114,7 @@ __C infiniStatus_t infiniopLPNorm(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroyLPNormDescriptor(infiniopLPNormDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                    \

src/infiniop/ops/mul/operator.cc

@@ -18,7 +18,7 @@
 #include "moore/mul_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreateMulDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateMulDescriptor(
     infiniopHandle_t handle,
     infiniopMulDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t c_desc,
@@ -68,7 +68,7 @@ __C infiniStatus_t infiniopCreateMulDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetMulWorkspaceSize(infiniopMulDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetMulWorkspaceSize(infiniopMulDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                               \
     case CASE:                                                                             \
@@ -109,7 +109,7 @@ __C infiniStatus_t infiniopGetMulWorkspaceSize(infiniopMulDescriptor_t desc, siz
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopMul(
+__INFINI_C infiniStatus_t infiniopMul(
     infiniopMulDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -157,7 +157,7 @@ __C infiniStatus_t infiniopMul(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroyMulDescriptor(infiniopMulDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                \

src/infiniop/ops/ones/operator.cc

@@ -15,7 +15,7 @@
 #include "moore/ones_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreateOnesDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateOnesDescriptor(
     infiniopHandle_t handle,
     infiniopOnesDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -60,7 +60,7 @@ __C infiniStatus_t infiniopCreateOnesDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetOnesWorkspaceSize(infiniopOnesDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetOnesWorkspaceSize(infiniopOnesDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                \
     case CASE:                                                                              \
@@ -98,7 +98,7 @@ __C infiniStatus_t infiniopGetOnesWorkspaceSize(infiniopOnesDescriptor_t desc, s
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopOnes(
+__INFINI_C infiniStatus_t infiniopOnes(
     infiniopOnesDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -142,7 +142,7 @@ __C infiniStatus_t infiniopOnes(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroyOnesDescriptor(infiniopOnesDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                 \

src/infiniop/ops/paged_attention/operator.cc

@@ -12,7 +12,7 @@
 #include "metax/paged_attention_metax.h"
 #endif
 
-__C infiniStatus_t infiniopCreatePagedAttentionDescriptor(
+__INFINI_C infiniStatus_t infiniopCreatePagedAttentionDescriptor(
     infiniopHandle_t handle,
     infiniopPagedAttentionDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t out_desc,
@@ -54,7 +54,7 @@ __C infiniStatus_t infiniopCreatePagedAttentionDescriptor(
     }
 }
 
-__C infiniStatus_t infiniopGetPagedAttentionWorkspaceSize(
+__INFINI_C infiniStatus_t infiniopGetPagedAttentionWorkspaceSize(
     infiniopPagedAttentionDescriptor_t desc,
     size_t *size) {
 
@@ -84,7 +84,7 @@ __C infiniStatus_t infiniopGetPagedAttentionWorkspaceSize(
     }
 }
 
-__C infiniStatus_t infiniopPagedAttention(
+__INFINI_C infiniStatus_t infiniopPagedAttention(
     infiniopPagedAttentionDescriptor_t desc,
     void *workspace, size_t workspace_size,
     void *out, const void *q, const void *k_cache, const void *v_cache,
@@ -118,7 +118,7 @@ __C infiniStatus_t infiniopPagedAttention(
     }
 }
 
-__C infiniStatus_t infiniopDestroyPagedAttentionDescriptor(
+__INFINI_C infiniStatus_t infiniopDestroyPagedAttentionDescriptor(
     infiniopPagedAttentionDescriptor_t desc) {
 
 #define DESTROY(CASE, NAMESPACE)                                                     \

src/infiniop/ops/paged_attention_prefill/operator.cc

@@ -12,7 +12,7 @@
 #include "moore/paged_attention_prefill_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreatePagedAttentionPrefillDescriptor(
+__INFINI_C infiniStatus_t infiniopCreatePagedAttentionPrefillDescriptor(
     infiniopHandle_t handle,
     infiniopPagedAttentionPrefillDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t out_desc,
@@ -56,7 +56,7 @@ __C infiniStatus_t infiniopCreatePagedAttentionPrefillDescriptor(
     }
 }
 
-__C infiniStatus_t infiniopGetPagedAttentionPrefillWorkspaceSize(
+__INFINI_C infiniStatus_t infiniopGetPagedAttentionPrefillWorkspaceSize(
     infiniopPagedAttentionPrefillDescriptor_t desc,
     size_t *size) {
 
@@ -86,7 +86,7 @@ __C infiniStatus_t infiniopGetPagedAttentionPrefillWorkspaceSize(
     }
 }
 
-__C infiniStatus_t infiniopPagedAttentionPrefill(
+__INFINI_C infiniStatus_t infiniopPagedAttentionPrefill(
     infiniopPagedAttentionPrefillDescriptor_t desc,
     void *workspace, size_t workspace_size,
     void *out, const void *q, const void *k_cache, const void *v_cache,
@@ -123,7 +123,7 @@ __C infiniStatus_t infiniopPagedAttentionPrefill(
     }
 }
 
-__C infiniStatus_t infiniopDestroyPagedAttentionPrefillDescriptor(
+__INFINI_C infiniStatus_t infiniopDestroyPagedAttentionPrefillDescriptor(
     infiniopPagedAttentionPrefillDescriptor_t desc) {
 
 #define DESTROY(CASE, NAMESPACE)                                                             \

src/infiniop/ops/paged_caching/operator.cc

@@ -12,7 +12,7 @@
 #include "moore/paged_caching_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreatePagedCachingDescriptor(
+__INFINI_C infiniStatus_t infiniopCreatePagedCachingDescriptor(
     infiniopHandle_t handle,
     infiniopPagedCachingDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t k_cache_desc,
@@ -49,7 +49,7 @@ __C infiniStatus_t infiniopCreatePagedCachingDescriptor(
     }
 }
 
-__C infiniStatus_t infiniopGetPagedCachingWorkspaceSize(
+__INFINI_C infiniStatus_t infiniopGetPagedCachingWorkspaceSize(
     infiniopPagedCachingDescriptor_t desc,
     size_t *size) {
 
@@ -79,7 +79,7 @@ __C infiniStatus_t infiniopGetPagedCachingWorkspaceSize(
     }
 }
 
-__C infiniStatus_t infiniopPagedCaching(
+__INFINI_C infiniStatus_t infiniopPagedCaching(
     infiniopPagedCachingDescriptor_t desc,
     void *workspace, size_t workspace_size,
     void *k_cache, void *v_cache,
@@ -113,7 +113,7 @@ __C infiniStatus_t infiniopPagedCaching(
     }
 }
 
-__C infiniStatus_t infiniopDestroyPagedCachingDescriptor(
+__INFINI_C infiniStatus_t infiniopDestroyPagedCachingDescriptor(
     infiniopPagedCachingDescriptor_t desc) {
 
 #define DESTROY(CASE, NAMESPACE)                                                   \

src/infiniop/ops/quant/per_channel_quant_int8/operator.cc

@@ -9,7 +9,7 @@
 #include "moore/per_channel_quant_int8_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreatePerChannelQuantI8Descriptor(infiniopHandle_t handle,
+__INFINI_C infiniStatus_t infiniopCreatePerChannelQuantI8Descriptor(infiniopHandle_t handle,
                                                              infiniopPerChannelQuantI8Descriptor_t *desc_ptr,
                                                              infiniopTensorDescriptor_t x_packed_desc,
                                                              infiniopTensorDescriptor_t x_scale_desc,
@@ -40,7 +40,7 @@ __C infiniStatus_t infiniopCreatePerChannelQuantI8Descriptor(infiniopHandle_t ha
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetPerChannelQuantI8WorkspaceSize(infiniopPerChannelQuantI8Descriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetPerChannelQuantI8WorkspaceSize(infiniopPerChannelQuantI8Descriptor_t desc, size_t *size) {
     switch (desc->device_type) {
 #define GET(CASE, NAMESPACE)                                                                                     \
     case CASE:                                                                                                   \
@@ -61,7 +61,7 @@ __C infiniStatus_t infiniopGetPerChannelQuantI8WorkspaceSize(infiniopPerChannelQ
 #undef GET
 }
 
-__C infiniStatus_t infiniopPerChannelQuantI8(infiniopPerChannelQuantI8Descriptor_t desc,
+__INFINI_C infiniStatus_t infiniopPerChannelQuantI8(infiniopPerChannelQuantI8Descriptor_t desc,
                                              void *workspace,
                                              size_t workspace_size,
                                              void *x_packed,
@@ -90,7 +90,7 @@ __C infiniStatus_t infiniopPerChannelQuantI8(infiniopPerChannelQuantI8Descriptor
 #undef QUANT
 }
 
-__C infiniStatus_t infiniopDestroyPerChannelQuantI8Descriptor(infiniopPerChannelQuantI8Descriptor_t desc) {
+__INFINI_C infiniStatus_t infiniopDestroyPerChannelQuantI8Descriptor(infiniopPerChannelQuantI8Descriptor_t desc) {
 #define DESTROY(CASE, NAMESPACE)                                                            \
     case CASE:                                                                              \
         delete reinterpret_cast<op::per_channel_quant_int8::NAMESPACE::Descriptor *>(desc); \

src/infiniop/ops/random_sample/operator.cc

@@ -24,7 +24,7 @@
 #include "kunlun/random_sample_kunlun.h"
 #endif
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopCreateRandomSampleDescriptor(
     infiniopHandle_t handle,
     infiniopRandomSampleDescriptor_t *desc_ptr,
@@ -82,7 +82,7 @@ infiniopCreateRandomSampleDescriptor(
 #undef CREATE
 };
 
-__C infiniStatus_t infiniopGetRandomSampleWorkspaceSize(
+__INFINI_C infiniStatus_t infiniopGetRandomSampleWorkspaceSize(
     infiniopRandomSampleDescriptor_t desc,
     size_t *size) {
 
@@ -136,7 +136,7 @@ __C infiniStatus_t infiniopGetRandomSampleWorkspaceSize(
 #undef GET
 }
 
-__C infiniStatus_t infiniopRandomSample(
+__INFINI_C infiniStatus_t infiniopRandomSample(
     infiniopRandomSampleDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -200,7 +200,7 @@ __C infiniStatus_t infiniopRandomSample(
 #undef CALCULATE
 }
 
-__C infiniStatus_t infiniopDestroyRandomSampleDescriptor(
+__INFINI_C infiniStatus_t infiniopDestroyRandomSampleDescriptor(
     infiniopRandomSampleDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                          \

src/infiniop/ops/rearrange/operator.cc

@@ -24,7 +24,7 @@
 #include "kunlun/rearrange_kunlun.h"
 #endif
 
-__C infiniStatus_t infiniopCreateRearrangeDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateRearrangeDescriptor(
     infiniopHandle_t handle,
     infiniopRearrangeDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t dst,
@@ -80,7 +80,7 @@ __C infiniStatus_t infiniopCreateRearrangeDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopRearrange(
+__INFINI_C infiniStatus_t infiniopRearrange(
     infiniopRearrangeDescriptor_t desc,
     void *dst,
     const void *src,
@@ -134,7 +134,7 @@ __C infiniStatus_t infiniopRearrange(
 #undef CALCULATE
 }
 
-__C infiniStatus_t infiniopDestroyRearrangeDescriptor(
+__INFINI_C infiniStatus_t infiniopDestroyRearrangeDescriptor(
     infiniopRearrangeDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                      \

src/infiniop/ops/relu/operator.cc

@@ -14,7 +14,7 @@
 #endif
 #endif
 
-__C infiniStatus_t infiniopCreateReluDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateReluDescriptor(
     infiniopHandle_t handle,
     infiniopReluDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -58,7 +58,7 @@ __C infiniStatus_t infiniopCreateReluDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetReluWorkspaceSize(infiniopReluDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetReluWorkspaceSize(infiniopReluDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                \
     case CASE:                                                                              \
@@ -95,7 +95,7 @@ __C infiniStatus_t infiniopGetReluWorkspaceSize(infiniopReluDescriptor_t desc, s
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopRelu(
+__INFINI_C infiniStatus_t infiniopRelu(
     infiniopReluDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -138,7 +138,7 @@ __C infiniStatus_t infiniopRelu(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroyReluDescriptor(infiniopReluDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                 \

src/infiniop/ops/rms_norm/operator.cc

@@ -24,7 +24,7 @@
 #include "kunlun/rms_norm_kunlun.h"
 #endif
 
-__C infiniStatus_t infiniopCreateRMSNormDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateRMSNormDescriptor(
     infiniopHandle_t handle,
     infiniopRMSNormDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -83,7 +83,7 @@ __C infiniStatus_t infiniopCreateRMSNormDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                    \
     case CASE:                                                                                  \
@@ -131,7 +131,7 @@ __C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t d
 #undef GET
 }
 
-__C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *workspace, size_t workspace_size,
+__INFINI_C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *workspace, size_t workspace_size,
                                    void *y, const void *x, const void *w, void *stream) {
 
 #define CALCULATE(CASE, NAMESPACE)                                                       \
@@ -180,7 +180,7 @@ __C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *works
 #undef CALCULATE
 }
 
-__C infiniStatus_t infiniopDestroyRMSNormDescriptor(infiniopRMSNormDescriptor_t desc) {
+__INFINI_C infiniStatus_t infiniopDestroyRMSNormDescriptor(infiniopRMSNormDescriptor_t desc) {
 
 #define DESTROY(CASE, NAMESPACE)                                              \
     case CASE:                                                                \

src/infiniop/ops/rope/operator.cc

@@ -24,7 +24,7 @@
 #include "moore/rope_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreateRoPEDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateRoPEDescriptor(
     infiniopHandle_t handle,
     infiniopRoPEDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y,
@@ -87,7 +87,7 @@ __C infiniStatus_t infiniopCreateRoPEDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetRoPEWorkspaceSize(infiniopRoPEDescriptor_t desc,
+__INFINI_C infiniStatus_t infiniopGetRoPEWorkspaceSize(infiniopRoPEDescriptor_t desc,
                                                 size_t *size) {
 #define GET(CASE, NAMESPACE)                                                                      \
     case CASE:                                                                                    \
@@ -135,7 +135,7 @@ __C infiniStatus_t infiniopGetRoPEWorkspaceSize(infiniopRoPEDescriptor_t desc,
 #undef GET
 }
 
-__C infiniStatus_t infiniopRoPE(
+__INFINI_C infiniStatus_t infiniopRoPE(
     infiniopRoPEDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -192,7 +192,7 @@ __C infiniStatus_t infiniopRoPE(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroyRoPEDescriptor(infiniopRoPEDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                 \

src/infiniop/ops/scaled_mm/operator.cc

@@ -10,7 +10,7 @@
 #include "moore/int8_gemm_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreateI8GemmDescriptor(infiniopHandle_t handle,
+__INFINI_C infiniStatus_t infiniopCreateI8GemmDescriptor(infiniopHandle_t handle,
                                                   infiniopI8GemmDescriptor_t *desc_ptr,
                                                   infiniopTensorDescriptor_t out_desc,
                                                   infiniopTensorDescriptor_t bias_desc,
@@ -45,7 +45,7 @@ __C infiniStatus_t infiniopCreateI8GemmDescriptor(infiniopHandle_t handle,
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetI8GemmWorkspaceSize(infiniopI8GemmDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetI8GemmWorkspaceSize(infiniopI8GemmDescriptor_t desc, size_t *size) {
     switch (desc->device_type) {
 #define GET(CASE, NAMESPACE)                                                                     \
     case CASE:                                                                                   \
@@ -66,7 +66,7 @@ __C infiniStatus_t infiniopGetI8GemmWorkspaceSize(infiniopI8GemmDescriptor_t des
 #undef GET
 }
 
-__C infiniStatus_t infiniopI8Gemm(infiniopI8GemmDescriptor_t desc,
+__INFINI_C infiniStatus_t infiniopI8Gemm(infiniopI8GemmDescriptor_t desc,
                                   void *workspace,
                                   size_t workspace_size,
                                   void *out,
@@ -96,7 +96,7 @@ __C infiniStatus_t infiniopI8Gemm(infiniopI8GemmDescriptor_t desc,
 #undef CACULATE
 }
 
-__C infiniStatus_t infiniopDestroyI8GemmDescriptor(infiniopI8GemmDescriptor_t desc) {
+__INFINI_C infiniStatus_t infiniopDestroyI8GemmDescriptor(infiniopI8GemmDescriptor_t desc) {
 #define DESTROY(CASE, NAMESPACE)                                            \
     case CASE:                                                              \
         delete reinterpret_cast<op::i8gemm::NAMESPACE::Descriptor *>(desc); \

src/infiniop/ops/sigmoid/operator.cc

@@ -9,7 +9,7 @@
 #include "nvidia/sigmoid_nvidia.cuh"
 #endif
 
-__C infiniStatus_t infiniopCreateSigmoidDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateSigmoidDescriptor(
     infiniopHandle_t handle,
     infiniopSigmoidDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -48,7 +48,7 @@ __C infiniStatus_t infiniopCreateSigmoidDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetSigmoidWorkspaceSize(infiniopSigmoidDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetSigmoidWorkspaceSize(infiniopSigmoidDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                   \
     case CASE:                                                                                 \
@@ -79,7 +79,7 @@ __C infiniStatus_t infiniopGetSigmoidWorkspaceSize(infiniopSigmoidDescriptor_t d
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopSigmoid(
+__INFINI_C infiniStatus_t infiniopSigmoid(
     infiniopSigmoidDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -116,7 +116,7 @@ __C infiniStatus_t infiniopSigmoid(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroySigmoidDescriptor(infiniopSigmoidDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                    \

src/infiniop/ops/silu/operator.cc

@@ -15,7 +15,7 @@
 #include "moore/silu_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreateSiluDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateSiluDescriptor(
     infiniopHandle_t handle,
     infiniopSiluDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t output_desc,
@@ -57,7 +57,7 @@ __C infiniStatus_t infiniopCreateSiluDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetSiluWorkspaceSize(infiniopSiluDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetSiluWorkspaceSize(infiniopSiluDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                \
     case CASE:                                                                              \
@@ -92,7 +92,7 @@ __C infiniStatus_t infiniopGetSiluWorkspaceSize(infiniopSiluDescriptor_t desc, s
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopSilu(
+__INFINI_C infiniStatus_t infiniopSilu(
     infiniopSiluDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -133,7 +133,7 @@ __C infiniStatus_t infiniopSilu(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroySiluDescriptor(infiniopSiluDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                 \

src/infiniop/ops/silu_and_mul/operator.cc

@@ -6,7 +6,7 @@
 #include "moore/silu_and_mul_moore.h"
 #endif
 
-__C infiniStatus_t infiniopCreateSiluAndMulDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateSiluAndMulDescriptor(
     infiniopHandle_t handle,
     infiniopSiluAndMulDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -31,7 +31,7 @@ __C infiniStatus_t infiniopCreateSiluAndMulDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetSiluAndMulWorkspaceSize(infiniopSiluAndMulDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetSiluAndMulWorkspaceSize(infiniopSiluAndMulDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                              \
     case CASE:                                                                                            \
@@ -49,7 +49,7 @@ __C infiniStatus_t infiniopGetSiluAndMulWorkspaceSize(infiniopSiluAndMulDescript
 #undef GET
 }
 
-__C infiniStatus_t infiniopSiluAndMul(
+__INFINI_C infiniStatus_t infiniopSiluAndMul(
     infiniopSiluAndMulDescriptor_t desc,
     void *workspace, size_t workspace_size,
     void *y,
@@ -72,7 +72,7 @@ __C infiniStatus_t infiniopSiluAndMul(
 #undef CALCULATE
 }
 
-__C infiniStatus_t infiniopDestroySiluAndMulDescriptor(infiniopSiluAndMulDescriptor_t desc) {
+__INFINI_C infiniStatus_t infiniopDestroySiluAndMulDescriptor(infiniopSiluAndMulDescriptor_t desc) {
 
 #define DESTROY(CASE, NAMESPACE)                                                        \
     case CASE:                                                                          \

src/infiniop/ops/softmax/operator.cc

@@ -6,7 +6,7 @@
 #include "nvidia/softmax_nvidia.cuh"
 #endif
 
-__C infiniStatus_t infiniopCreateSoftmaxDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateSoftmaxDescriptor(
     infiniopHandle_t handle,
     infiniopSoftmaxDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -42,7 +42,7 @@ __C infiniStatus_t infiniopCreateSoftmaxDescriptor(
     }
 }
 
-__C infiniStatus_t infiniopGetSoftmaxWorkspaceSize(infiniopSoftmaxDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetSoftmaxWorkspaceSize(infiniopSoftmaxDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                   \
     case CASE:                                                                                 \
@@ -70,7 +70,7 @@ __C infiniStatus_t infiniopGetSoftmaxWorkspaceSize(infiniopSoftmaxDescriptor_t d
     }
 }
 
-__C infiniStatus_t infiniopSoftmax(
+__INFINI_C infiniStatus_t infiniopSoftmax(
     infiniopSoftmaxDescriptor_t desc,
     void *workspace, size_t workspace_size,
     void *y,
@@ -103,7 +103,7 @@ __C infiniStatus_t infiniopSoftmax(
     }
 }
 
-__C infiniStatus_t infiniopDestroySoftmaxDescriptor(infiniopSoftmaxDescriptor_t desc) {
+__INFINI_C infiniStatus_t infiniopDestroySoftmaxDescriptor(infiniopSoftmaxDescriptor_t desc) {
 
 #define DESTROY(CASE, NAMESPACE)                                             \
     case CASE:                                                               \

src/infiniop/ops/softplus/operator.cc

@@ -15,7 +15,7 @@
 #include "kunlun/softplus_kunlun.h"
 #endif
 
-__C infiniStatus_t infiniopCreateSoftplusDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateSoftplusDescriptor(
     infiniopHandle_t handle,
     infiniopSoftplusDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t y_desc,
@@ -60,7 +60,7 @@ __C infiniStatus_t infiniopCreateSoftplusDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetSoftplusWorkspaceSize(infiniopSoftplusDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetSoftplusWorkspaceSize(infiniopSoftplusDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                                    \
     case CASE:                                                                                  \
@@ -98,7 +98,7 @@ __C infiniStatus_t infiniopGetSoftplusWorkspaceSize(infiniopSoftplusDescriptor_t
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopSoftplus(
+__INFINI_C infiniStatus_t infiniopSoftplus(
     infiniopSoftplusDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -142,7 +142,7 @@ __C infiniStatus_t infiniopSoftplus(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroySoftplusDescriptor(infiniopSoftplusDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                     \

src/infiniop/ops/sub/operator.cc

@@ -15,7 +15,7 @@
 #include "kunlun/sub_kunlun.h"
 #endif
 
-__C infiniStatus_t infiniopCreateSubDescriptor(
+__INFINI_C infiniStatus_t infiniopCreateSubDescriptor(
     infiniopHandle_t handle,
     infiniopSubDescriptor_t *desc_ptr,
     infiniopTensorDescriptor_t c_desc,
@@ -62,7 +62,7 @@ __C infiniStatus_t infiniopCreateSubDescriptor(
 #undef CREATE
 }
 
-__C infiniStatus_t infiniopGetSubWorkspaceSize(infiniopSubDescriptor_t desc, size_t *size) {
+__INFINI_C infiniStatus_t infiniopGetSubWorkspaceSize(infiniopSubDescriptor_t desc, size_t *size) {
 
 #define GET(CASE, NAMESPACE)                                                               \
     case CASE:                                                                             \
@@ -100,7 +100,7 @@ __C infiniStatus_t infiniopGetSubWorkspaceSize(infiniopSubDescriptor_t desc, siz
     return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
 }
 
-__C infiniStatus_t infiniopSub(
+__INFINI_C infiniStatus_t infiniopSub(
     infiniopSubDescriptor_t desc,
     void *workspace,
     size_t workspace_size,
@@ -145,7 +145,7 @@ __C infiniStatus_t infiniopSub(
 #undef CALCULATE
 }
 
-__C infiniStatus_t
+__INFINI_C infiniStatus_t
 infiniopDestroySubDescriptor(infiniopSubDescriptor_t desc) {
 
 #define DELETE(CASE, NAMESPACE)                                                \