Skip to content

MegEngine
MegEngine

CHEL01. / MegEngine
与 Fork 源项目一致

Fork自 MegEngine 天元 / MegEngine

通知 1
Star 0
Fork 0
MegEngine
MegEngine

体验新版 GitCode，发现更多精彩内容 >>

提交 eaa18018 编写于 作者: M Megvii Engine Team
浏览文件
操作

feat(x86/rvv): opt gi intrinsic helper 

for rvv, detail: https://github.com/riscv-collab/riscv-gnu-toolchain/issues/1106

GitOrigin-RevId: 27615584c0a502321569f79d13dcc88aff65d253
上级 399db31a
隐藏空白更改
内联 并排
Showing with 156 addition and 101 deletion
dnn/src/fallback/conv_bias/gi/intrinsic_helper.h
浏览文件 @ eaa18018
......@@ -258,189 +258,244 @@ struct StoreOcxOw8Remain {
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 0, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][6], c[0][7]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][2], c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][4], c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
op(vis(c[1][6], c[1][7]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 24));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
op(vis(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[0][7]), reinterpret_cast<T3>(dst_ptr + 28));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
op(vis(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 12));
op(vis(c[1][4]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
op(vis(c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 20));
op(vis(c[1][6]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 24));
op(vis(c[1][7]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 28));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 8, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][6], c[0][7]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][2], c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][4], c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
op(vis(c[1][6], c[1][7]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 24));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
op(vis(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[0][7]), reinterpret_cast<T3>(dst_ptr + 28));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
op(vis(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 12));
op(vis(c[1][4]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
op(vis(c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 20));
op(vis(c[1][6]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 24));
op(vis(c[1][7]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 28));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 7, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitor<typename Op::src_ctype> vis0;
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis0(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][2], c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][4], c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
op(vis0(c[1][6]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 24));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
op(vis(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
op(vis(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 12));
op(vis(c[1][4]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
op(vis(c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 20));
op(vis(c[1][6]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 24));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 6, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][2], c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][4], c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
op(vis(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 12));
op(vis(c[1][4]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
op(vis(c[1][5]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 20));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 5, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitor<typename Op::src_ctype> vis0;
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis0(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][2], c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis0(c[1][4]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
op(vis(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 12));
op(vis(c[1][4]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 16));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 4, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][2], c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
op(vis(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
op(vis(c[1][3]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 12));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 3, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitor<typename Op::src_ctype> vis0;
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis0(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis0(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
op(vis(c[1][2]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 8));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 2, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[1][0], c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
op(vis(c[1][1]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc + 4));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<2, 1, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int ld_dst_oc) {
ParamElemFixLenVisitor<typename Op::src_ctype> vis0;
op(vis0(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis0(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[1][0]), reinterpret_cast<T3>(dst_ptr + ld_dst_oc));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 0, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][6], c[0][7]), reinterpret_cast<T3>(dst_ptr + 24));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
op(vis(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[0][7]), reinterpret_cast<T3>(dst_ptr + 28));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 8, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][6], c[0][7]), reinterpret_cast<T3>(dst_ptr + 24));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
op(vis(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
op(vis(c[0][7]), reinterpret_cast<T3>(dst_ptr + 28));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 7, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitor<typename Op::src_ctype> vis0;
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis0(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
op(vis(c[0][6]), reinterpret_cast<T3>(dst_ptr + 24));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 6, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][4], c[0][5]), reinterpret_cast<T3>(dst_ptr + 16));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
op(vis(c[0][5]), reinterpret_cast<T3>(dst_ptr + 20));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 5, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitor<typename Op::src_ctype> vis0;
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis0(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
op(vis(c[0][4]), reinterpret_cast<T3>(dst_ptr + 16));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 4, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][2], c[0][3]), reinterpret_cast<T3>(dst_ptr + 8));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
op(vis(c[0][3]), reinterpret_cast<T3>(dst_ptr + 12));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 3, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitor<typename Op::src_ctype> vis0;
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
op(vis0(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
op(vis(c[0][2]), reinterpret_cast<T3>(dst_ptr + 8));
}
};
template <typename Op, typename T, typename T2, typename T3>
struct StoreOcxOw8Remain<1, 2, Op, T, T2, T3> {
static GI_FORCEINLINE void impl(T& c, const Op& op, T2 dst_ptr, int) {
ParamElemFixLenVisitorV2<typename Op::src_ctype> vis;
op(vis(c[0][0], c[0][1]), reinterpret_cast<T3>(dst_ptr));
ParamElemFixLenVisitor<typename Op::src_ctype> vis;
op(vis(c[0][0]), reinterpret_cast<T3>(dst_ptr));
op(vis(c[0][1]), reinterpret_cast<T3>(dst_ptr + 4));
}
};
template <typename Op, typename T, typename T2, typename T3>
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或