[T1-1-2] wawahejun

include/infiniop.h

@@ -3,19 +3,37 @@
 
 #include "infiniop/handle.h"
 #include "infiniop/ops/add.h"
+#include "infiniop/ops/and.h"
 #include "infiniop/ops/attention.h"
+#include "infiniop/ops/cast.h"
 #include "infiniop/ops/causal_softmax.h"
 #include "infiniop/ops/clip.h"
 #include "infiniop/ops/conv.h"
+#include "infiniop/ops/cos.h"
+#include "infiniop/ops/crossentropyloss_backward.h"
+#include "infiniop/ops/div.h"
+#include "infiniop/ops/equal.h"
+#include "infiniop/ops/exp.h"
+#include "infiniop/ops/gelu.h"
+#include "infiniop/ops/gelu_backward.h"
 #include "infiniop/ops/gemm.h"
+#include "infiniop/ops/hardswish.h"
+#include "infiniop/ops/leaky_relu.h"
 #include "infiniop/ops/mul.h"
+#include "infiniop/ops/or.h"
 #include "infiniop/ops/random_sample.h"
 #include "infiniop/ops/rearrange.h"
 #include "infiniop/ops/relu.h"
+#include "infiniop/ops/relu_backward.h"
 #include "infiniop/ops/rms_norm.h"
 #include "infiniop/ops/rope.h"
+#include "infiniop/ops/sigmoid_backward.h"
+#include "infiniop/ops/silu.h"
+#include "infiniop/ops/sin.h"
 #include "infiniop/ops/sub.h"
 #include "infiniop/ops/swiglu.h"
+#include "infiniop/ops/tanh.h"
+#include "infiniop/ops/where.h"
 #include "infiniop/tensor_descriptor.h"
 
 #endif // __INFINIOP_API_H__

include/infiniop/ops/and.h

@@ -0,0 +1,26 @@
+#ifndef __INFINIOP_AND_API_H__
+#define __INFINIOP_AND_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopAndDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateAndDescriptor(infiniopHandle_t handle,
+                                                        infiniopAndDescriptor_t *desc_ptr,
+                                                        infiniopTensorDescriptor_t c,
+                                                        infiniopTensorDescriptor_t a,
+                                                        infiniopTensorDescriptor_t b);
+
+__C __export infiniStatus_t infiniopGetAndWorkspaceSize(infiniopAndDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopAnd(infiniopAndDescriptor_t desc,
+                                        void *workspace,
+                                        size_t workspace_size,
+                                        void *c,
+                                        const void *a,
+                                        const void *b,
+                                        void *stream);
+
+__C __export infiniStatus_t infiniopDestroyAndDescriptor(infiniopAndDescriptor_t desc);
+
+#endif

include/infiniop/ops/cast.h

@@ -0,0 +1,24 @@
+#ifndef __INFINIOP_CAST_API_H__
+#define __INFINIOP_CAST_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopCastDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateCastDescriptor(infiniopHandle_t handle,
+                                                        infiniopCastDescriptor_t *desc_ptr,
+                                                        infiniopTensorDescriptor_t output,
+                                                        infiniopTensorDescriptor_t input);
+
+__C __export infiniStatus_t infiniopGetCastWorkspaceSize(infiniopCastDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopCast(infiniopCastDescriptor_t desc,
+                                        void *workspace,
+                                        size_t workspace_size,
+                                        void *output,
+                                        const void *input,
+                                        void *stream);
+
+__C __export infiniStatus_t infiniopDestroyCastDescriptor(infiniopCastDescriptor_t desc);
+
+#endif

include/infiniop/ops/cos.h

@@ -0,0 +1,24 @@
+#ifndef __INFINIOP_COS_API_H__
+#define __INFINIOP_COS_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopCosDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateCosDescriptor(infiniopHandle_t handle,
+                                                        infiniopCosDescriptor_t *desc_ptr,
+                                                        infiniopTensorDescriptor_t output,
+                                                        infiniopTensorDescriptor_t input);
+
+__C __export infiniStatus_t infiniopGetCosWorkspaceSize(infiniopCosDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopCos(infiniopCosDescriptor_t desc,
+                                        void *workspace,
+                                        size_t workspace_size,
+                                        void *output,
+                                        const void *input,
+                                        void *stream);
+
+__C __export infiniStatus_t infiniopDestroyCosDescriptor(infiniopCosDescriptor_t desc);
+
+#endif

include/infiniop/ops/crossentropyloss_backward.h

@@ -0,0 +1,26 @@
+#ifndef __INFINIOP_CROSSENTROPYLOSS_BACKWARD_API_H__
+#define __INFINIOP_CROSSENTROPYLOSS_BACKWARD_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopCrossEntropyLossBackwardDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateCrossEntropyLossBackwardDescriptor(infiniopHandle_t handle,
+                                                                             infiniopCrossEntropyLossBackwardDescriptor_t *desc_ptr,
+                                                                             infiniopTensorDescriptor_t grad_logits,
+                                                                             infiniopTensorDescriptor_t probs,
+                                                                             infiniopTensorDescriptor_t target);
+
+__C __export infiniStatus_t infiniopGetCrossEntropyLossBackwardWorkspaceSize(infiniopCrossEntropyLossBackwardDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopCrossEntropyLossBackward(infiniopCrossEntropyLossBackwardDescriptor_t desc,
+                                                             void *workspace,
+                                                             size_t workspace_size,
+                                                             void *grad_logits,
+                                                             const void *probs,
+                                                             const void *target,
+                                                             void *stream);
+
+__C __export infiniStatus_t infiniopDestroyCrossEntropyLossBackwardDescriptor(infiniopCrossEntropyLossBackwardDescriptor_t desc);
+
+#endif

include/infiniop/ops/div.h

@@ -0,0 +1,26 @@
+#ifndef __INFINIOP_DIV_API_H__
+#define __INFINIOP_DIV_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDivDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateDivDescriptor(infiniopHandle_t handle,
+                                                        infiniopDivDescriptor_t *desc_ptr,
+                                                        infiniopTensorDescriptor_t c,
+                                                        infiniopTensorDescriptor_t a,
+                                                        infiniopTensorDescriptor_t b);
+
+__C __export infiniStatus_t infiniopGetDivWorkspaceSize(infiniopDivDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopDiv(infiniopDivDescriptor_t desc,
+                                        void *workspace,
+                                        size_t workspace_size,
+                                        void *c,
+                                        const void *a,
+                                        const void *b,
+                                        void *stream);
+
+__C __export infiniStatus_t infiniopDestroyDivDescriptor(infiniopDivDescriptor_t desc);
+
+#endif

include/infiniop/ops/equal.h

@@ -0,0 +1,26 @@
+#ifndef __INFINIOP_EQUAL_API_H__
+#define __INFINIOP_EQUAL_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopEqualDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateEqualDescriptor(infiniopHandle_t handle,
+                                                          infiniopEqualDescriptor_t *desc_ptr,
+                                                          infiniopTensorDescriptor_t c,
+                                                          infiniopTensorDescriptor_t a,
+                                                          infiniopTensorDescriptor_t b);
+
+__C __export infiniStatus_t infiniopGetEqualWorkspaceSize(infiniopEqualDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopEqual(infiniopEqualDescriptor_t desc,
+                                          void *workspace,
+                                          size_t workspace_size,
+                                          void *c,
+                                          const void *a,
+                                          const void *b,
+                                          void *stream);
+
+__C __export infiniStatus_t infiniopDestroyEqualDescriptor(infiniopEqualDescriptor_t desc);
+
+#endif

include/infiniop/ops/exp.h

@@ -0,0 +1,24 @@
+#ifndef __INFINIOP_EXP_API_H__
+#define __INFINIOP_EXP_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopExpDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateExpDescriptor(infiniopHandle_t handle,
+                                                        infiniopExpDescriptor_t *desc_ptr,
+                                                        infiniopTensorDescriptor_t output,
+                                                        infiniopTensorDescriptor_t input);
+
+__C __export infiniStatus_t infiniopGetExpWorkspaceSize(infiniopExpDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopExp(infiniopExpDescriptor_t desc,
+                                        void *workspace,
+                                        size_t workspace_size,
+                                        void *output,
+                                        const void *input,
+                                        void *stream);
+
+__C __export infiniStatus_t infiniopDestroyExpDescriptor(infiniopExpDescriptor_t desc);
+
+#endif

include/infiniop/ops/gelu.h

@@ -0,0 +1,24 @@
+#ifndef __INFINIOP_GELU_API_H__
+#define __INFINIOP_GELU_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopGeluDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateGeluDescriptor(infiniopHandle_t handle,
+                                                         infiniopGeluDescriptor_t *desc_ptr,
+                                                         infiniopTensorDescriptor_t output,
+                                                         infiniopTensorDescriptor_t input);
+
+__C __export infiniStatus_t infiniopGetGeluWorkspaceSize(infiniopGeluDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopGelu(infiniopGeluDescriptor_t desc,
+                                         void *workspace,
+                                         size_t workspace_size,
+                                         void *output,
+                                         const void *input,
+                                         void *stream);
+
+__C __export infiniStatus_t infiniopDestroyGeluDescriptor(infiniopGeluDescriptor_t desc);
+
+#endif

include/infiniop/ops/gelu_backward.h

@@ -0,0 +1,26 @@
+#ifndef __INFINIOP_GELU_BACKWARD_API_H__
+#define __INFINIOP_GELU_BACKWARD_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopGeluBackwardDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateGeluBackwardDescriptor(infiniopHandle_t handle,
+                                                                 infiniopGeluBackwardDescriptor_t *desc_ptr,
+                                                                 infiniopTensorDescriptor_t grad_input,
+                                                                 infiniopTensorDescriptor_t input,
+                                                                 infiniopTensorDescriptor_t grad_output);
+
+__C __export infiniStatus_t infiniopGetGeluBackwardWorkspaceSize(infiniopGeluBackwardDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopGeluBackward(infiniopGeluBackwardDescriptor_t desc,
+                                                 void *workspace,
+                                                 size_t workspace_size,
+                                                 void *grad_input,
+                                                 const void *input,
+                                                 const void *grad_output,
+                                                 void *stream);
+
+__C __export infiniStatus_t infiniopDestroyGeluBackwardDescriptor(infiniopGeluBackwardDescriptor_t desc);
+
+#endif

include/infiniop/ops/hardswish.h

@@ -0,0 +1,24 @@
+#ifndef __INFINIOP_HARDSWISH_API_H__
+#define __INFINIOP_HARDSWISH_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopHardSwishDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateHardSwishDescriptor(infiniopHandle_t handle,
+                                                        infiniopHardSwishDescriptor_t *desc_ptr,
+                                                        infiniopTensorDescriptor_t output,
+                                                        infiniopTensorDescriptor_t input);
+
+__C __export infiniStatus_t infiniopGetHardSwishWorkspaceSize(infiniopHardSwishDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopHardSwish(infiniopHardSwishDescriptor_t desc,
+                                        void *workspace,
+                                        size_t workspace_size,
+                                        void *output,
+                                        const void *input,
+                                        void *stream);
+
+__C __export infiniStatus_t infiniopDestroyHardSwishDescriptor(infiniopHardSwishDescriptor_t desc);
+
+#endif

include/infiniop/ops/leaky_relu.h

@@ -0,0 +1,25 @@
+#ifndef __INFINIOP_LEAKY_RELU_API_H__
+#define __INFINIOP_LEAKY_RELU_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopLeakyReLUDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateLeakyReLUDescriptor(infiniopHandle_t handle,
+                                                        infiniopLeakyReLUDescriptor_t *desc_ptr,
+                                                        infiniopTensorDescriptor_t output,
+                                                        infiniopTensorDescriptor_t input,
+                                                        float negative_slope);
+
+__C __export infiniStatus_t infiniopGetLeakyReLUWorkspaceSize(infiniopLeakyReLUDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopLeakyReLU(infiniopLeakyReLUDescriptor_t desc,
+                                        void *workspace,
+                                        size_t workspace_size,
+                                        void *output,
+                                        const void *input,
+                                        void *stream);
+
+__C __export infiniStatus_t infiniopDestroyLeakyReLUDescriptor(infiniopLeakyReLUDescriptor_t desc);
+
+#endif

include/infiniop/ops/or.h

@@ -0,0 +1,26 @@
+#ifndef __INFINIOP_OR_API_H__
+#define __INFINIOP_OR_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopOrDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateOrDescriptor(infiniopHandle_t handle,
+                                                       infiniopOrDescriptor_t *desc_ptr,
+                                                       infiniopTensorDescriptor_t c,
+                                                       infiniopTensorDescriptor_t a,
+                                                       infiniopTensorDescriptor_t b);
+
+__C __export infiniStatus_t infiniopGetOrWorkspaceSize(infiniopOrDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopOr(infiniopOrDescriptor_t desc,
+                                       void *workspace,
+                                       size_t workspace_size,
+                                       void *c,
+                                       const void *a,
+                                       const void *b,
+                                       void *stream);
+
+__C __export infiniStatus_t infiniopDestroyOrDescriptor(infiniopOrDescriptor_t desc);
+
+#endif

include/infiniop/ops/relu_backward.h

@@ -0,0 +1,26 @@
+#ifndef __INFINIOP_RELU_BACKWARD_API_H__
+#define __INFINIOP_RELU_BACKWARD_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopReluBackwardDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateReluBackwardDescriptor(infiniopHandle_t handle,
+                                                                 infiniopReluBackwardDescriptor_t *desc_ptr,
+                                                                 infiniopTensorDescriptor_t grad_input,
+                                                                 infiniopTensorDescriptor_t input,
+                                                                 infiniopTensorDescriptor_t grad_output);
+
+__C __export infiniStatus_t infiniopGetReluBackwardWorkspaceSize(infiniopReluBackwardDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopReluBackward(infiniopReluBackwardDescriptor_t desc,
+                                                 void *workspace,
+                                                 size_t workspace_size,
+                                                 void *grad_input,
+                                                 const void *input,
+                                                 const void *grad_output,
+                                                 void *stream);
+
+__C __export infiniStatus_t infiniopDestroyReluBackwardDescriptor(infiniopReluBackwardDescriptor_t desc);
+
+#endif