Optimize mul_add for rvv(performance boost 15% ~ 32%)

kendryte · Sep 13, 2024 · 3371dd2 · 3371dd2
1 parent 7f3bbff
commit 3371dd2
Show file tree

Hide file tree

Showing 2 changed files with 23 additions and 2 deletions.
diff --git a/src/Native/include/nncase/ntt/arch/riscv64/primitive_ops.h b/src/Native/include/nncase/ntt/arch/riscv64/primitive_ops.h
@@ -969,6 +969,7 @@ REGISTER_RVV_KERNEL(INNER_PRODUCT_FLOAT32)
 REGISTER_RVV_INNER_PRODUCT_OP(float, inner_product_float32)
 
 // register mul_add kernel
+#if 0
 #define MUL_ADD_FLOAT32(lmul, mlen)                                            \
     inline vfloat32m##lmul##_t mul_add_float32(                                \
         const vfloat32m##lmul##_t &v1, const vfloat32m##lmul##_t &v2,          \
@@ -987,6 +988,26 @@ REGISTER_RVV_INNER_PRODUCT_OP(float, inner_product_float32)
         const vfloat32m##lmul##_t &v3, const size_t vl) {                      \
         return __riscv_vfmadd_vf_f32m##lmul(v2, s1, v3, vl);                   \
     }
+#else
+#define MUL_ADD_FLOAT32(lmul, mlen)                                            \
+    inline vfloat32m##lmul##_t mul_add_float32(                                \
+        const vfloat32m##lmul##_t &v1, const vfloat32m##lmul##_t &v2,          \
+        const vfloat32m##lmul##_t &v3, const size_t vl) {                      \
+        return __riscv_vfmacc_vv_f32m##lmul(v3, v1, v2, vl);                   \
+    }                                                                          \
+                                                                               \
+    inline vfloat32m##lmul##_t mul_add_float32(                                \
+        const vfloat32m##lmul##_t &v1, const float &s2,                        \
+        const vfloat32m##lmul##_t &v3, const size_t vl) {                      \
+        return __riscv_vfmacc_vf_f32m##lmul(v3, s2, v1, vl);                   \
+    }                                                                          \
+                                                                               \
+    inline vfloat32m##lmul##_t mul_add_float32(                                \
+        const float &s1, const vfloat32m##lmul##_t &v2,                        \
+        const vfloat32m##lmul##_t &v3, const size_t vl) {                      \
+        return __riscv_vfmacc_vf_f32m##lmul(v3, s1, v2, vl);                   \
+    }
+#endif
 
 REGISTER_RVV_KERNEL(MUL_ADD_FLOAT32)
 

diff --git a/src/Native/test/benchmark_test/benchmark_ntt.py b/src/Native/test/benchmark_test/benchmark_ntt.py
@@ -224,8 +224,8 @@ def __init__(self, target: str, bin_path: str):
                                         'tanh': '197',
                                         },
                               'matmul': {'no_pack': '92058',
-                                         'pack_K': '155648',
-                                         'pack_M': '75674',
+                                         'pack_K': '163840',
+                                         'pack_M': '76170',
                                          'pack_N': '92058',
                                          'pack_M_N': '47322',
                                          'pack_M_K': '95744',