34 lines
1.3 KiB
C
34 lines
1.3 KiB
C
#ifndef HTP_FUSION_H
|
|
#define HTP_FUSION_H
|
|
|
|
#include "htp-common.h"
|
|
|
|
enum htp_fusion_flags {
|
|
HTP_FUSE_ALLREDUCE_ADD = (1 << 1), // 2
|
|
HTP_FUSE_RMS_NORM_MUL = (1 << 2), // 4
|
|
HTP_FUSE_MUL_MAT_ADD = (1 << 3), // 8
|
|
HTP_FUSE_MUL_MAT_NX = (1 << 4), // 16
|
|
HTP_FUSE_MUL_MAT_ID_NX = (1 << 5), // 32
|
|
HTP_FUSE_GDN_CPY = (1 << 6), // 64
|
|
HTP_FUSE_REPEAT_GET_ROWS = (1 << 7), // 128
|
|
HTP_FUSE_UNARY_MUL_GLU = (1 << 8), // 256
|
|
HTP_FUSE_MUL_MUL_SCALAR = (1 << 9), // 512
|
|
HTP_FUSE_BINARY_NX = (1 << 10), // 1024
|
|
HTP_FUSE_RMS_NORM_SCALE = (1 << 11), // 2048
|
|
HTP_FUSE_SSM_CONV_SILU = (1 << 12), // 4096
|
|
};
|
|
|
|
static inline bool htp_fusion_enabled(int flag) {
|
|
if (opt_opfusion <= 0) return false;
|
|
if (opt_opfusion == 1) return true; // 1 enables all
|
|
return (opt_opfusion & flag) != 0;
|
|
}
|
|
|
|
bool htp_fusion_can_stack(const struct ggml_tensor * a);
|
|
bool htp_fusion_can_stack(const struct ggml_tensor * a, const struct ggml_tensor * b);
|
|
int htp_fusion_max(const struct ggml_cgraph * graph, int idx);
|
|
void htp_fusion_tag_graph(struct ggml_cgraph * graph);
|
|
bool htp_fusion_try_fuse(const struct ggml_hexagon_session * sess, struct htp_opnode & prev, const struct htp_opnode & cur);
|
|
|
|
#endif // HTP_FUSION_H
|