diff options
Diffstat (limited to 'inference-engine/thirdparty/mkl-dnn/src/cpu/cpu_reorder.cpp')
-rw-r--r-- | inference-engine/thirdparty/mkl-dnn/src/cpu/cpu_reorder.cpp | 378 |
1 files changed, 126 insertions, 252 deletions
diff --git a/inference-engine/thirdparty/mkl-dnn/src/cpu/cpu_reorder.cpp b/inference-engine/thirdparty/mkl-dnn/src/cpu/cpu_reorder.cpp index fbba44e71..ad791e6ff 100644 --- a/inference-engine/thirdparty/mkl-dnn/src/cpu/cpu_reorder.cpp +++ b/inference-engine/thirdparty/mkl-dnn/src/cpu/cpu_reorder.cpp @@ -34,261 +34,135 @@ namespace { using namespace mkldnn::impl::data_type; using namespace mkldnn::impl::memory_format; +#define REG_SR(idt, ifmt, odt, ofmt, ...) \ + simple_reorder_t<idt, ifmt, odt, ofmt, __VA_ARGS__>::pd_t::create + +#define REG_SR_BIDIR(idt, ifmt, odt, ofmt) \ + REG_SR(idt, ifmt, odt, ofmt, fmt_order::keep), \ + REG_SR(idt, ifmt, odt, ofmt, fmt_order::reverse) + +#define REG_SR_DIRECT_COPY(idt, odt) \ + REG_SR(idt, any, odt, any, fmt_order::any, spec::direct_copy), \ + REG_SR(idt, any, odt, any, fmt_order::any, spec::direct_copy_except_dim_0) + static const rpd_create_f cpu_reorder_impl_list[] = { - /* fp32 <-> fp32 */ - simple_reorder_t<f32, any, f32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<f32, any, f32, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, + /* winograd */ + wino_reorder_t<f32, f32>::pd_t::create, + wino_reorder_t<f32, s8>::pd_t::create, + +#ifdef __INTEL_COMPILER + /* direct copy for icc, which is faster than jitted code */ + REG_SR_DIRECT_COPY(f32, f32), + REG_SR_DIRECT_COPY(f32, s32), + REG_SR_DIRECT_COPY(f32, s8), +// REG_SR_DIRECT_COPY(f32, u8), FIXME: Disabled due to accuracy failure on int8 network + REG_SR_DIRECT_COPY(s32, f32), + REG_SR_DIRECT_COPY(s32, s32), + REG_SR_DIRECT_COPY(s32, s8), + REG_SR_DIRECT_COPY(s32, u8), + REG_SR_DIRECT_COPY(s8, f32), + REG_SR_DIRECT_COPY(s8, s32), + REG_SR_DIRECT_COPY(s8, s8), + REG_SR_DIRECT_COPY(s8, u8), + REG_SR_DIRECT_COPY(u8, f32), + REG_SR_DIRECT_COPY(u8, s32), + REG_SR_DIRECT_COPY(u8, s8), + REG_SR_DIRECT_COPY(u8, u8), +#endif + + /* jit */ jit_uni_reorder_create, - /* fp32 flat <-> fp32 blocked with tail */ - simple_reorder_t<f32, any, f32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, nCdhw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, nCdhw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, Oihw16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, Oihw16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, Ohwi16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, Ohwi16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, Oidhw16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, Oidhw16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, Odhwi16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, Odhwi16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, OIhw16o16i, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, OIhw16o16i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, OIhw16i16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, OIhw16i16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, OIdhw16o16i, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, OIdhw16o16i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, OIdhw16i16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, OIdhw16i16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, IOhw16o16i, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, IOhw16o16i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOihw16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOihw16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOhwi16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOhwi16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOidhw16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOidhw16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOdhwi16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOdhwi16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIhw16o16i, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIhw16o16i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIhw16i16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIhw16i16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIdhw16o16i, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIdhw16o16i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIdhw16i16o, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gOIdhw16i16o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, gIOhw16o16i, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, f32, gIOhw16o16i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nchw, f32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nchw, f32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nchw, f32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nchw, f32, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nchw, f32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nchw, f32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nChw8c, f32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nChw8c, f32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, any, f32, any, fmt_order::any, spec::reference>::pd_t::create, - /* reorder with quantization */ - wino_reorder_t<f32, goihw, s8, wino_fmt, fmt_order::keep>::pd_t::create, - wino_reorder_t<f32, oihw, s8, wino_fmt, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, any, s32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<f32, any, s8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<f32, any, u8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s32, any, f32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s32, any, s32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s32, any, s8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s32, any, u8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s8, any, f32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s8, any, s32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s8, any, s8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s8, any, u8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<u8, any, f32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<u8, any, s32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<u8, any, s8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<u8, any, u8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - /* nchw <-> nhwc */ - simple_reorder_t<f32, nchw, s32, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nchw, s32, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nchw, s8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nchw, s8, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nchw, u8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nchw, u8, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nchw, f32, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nchw, f32, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nchw, f32, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nchw, s32, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nchw, s8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nchw, s8, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nchw, u8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nchw, u8, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nchw, f32, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nchw, f32, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nchw, s32, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nchw, s32, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nchw, s8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nchw, s8, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nchw, u8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nchw, u8, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nchw, f32, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nchw, f32, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nchw, s32, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nchw, s32, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nchw, s8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nchw, s8, nhwc, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nchw, u8, nhwc, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nchw, u8, nhwc, fmt_order::reverse>::pd_t::create, - /* nhwc <-> nChw8c */ - simple_reorder_t<s8, nhwc, f32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, f32, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nhwc, s32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, s32, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nhwc, s8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, s8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nhwc, u8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, u8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nhwc, s32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nhwc, s32, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nhwc, s8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nhwc, s8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nhwc, u8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nhwc, u8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, s8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, s8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, s32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, s32, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, f32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, f32, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, u8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, u8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, s8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, s8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, s32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, s32, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, u8, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, u8, nChw8c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, f32, nChw8c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, f32, nChw8c, fmt_order::reverse>::pd_t::create, - /* nhwc <-> nChw16c */ - simple_reorder_t<s8, nhwc, f32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, f32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nhwc, s32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, s32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nhwc, s8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, s8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, nhwc, u8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, nhwc, u8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nhwc, s32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nhwc, s32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nhwc, s8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nhwc, s8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<f32, nhwc, u8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<f32, nhwc, u8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, s8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, s8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, s32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, s32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, f32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, f32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, f32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, nhwc, u8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, nhwc, u8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, s8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, s8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, s32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, s32, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, u8, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, u8, nChw16c, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, nhwc, f32, nChw16c, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, nhwc, f32, nChw16c, fmt_order::reverse>::pd_t::create, - /* s32 <-> fp32 */ - simple_reorder_t<f32, any, s32, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s32, any, f32, any, fmt_order::any, spec::reference>::pd_t::create, - /* s16 <-> fp32 */ - simple_reorder_t<f32, any, s16, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s16, any, f32, any, fmt_order::any, spec::reference>::pd_t::create, - /* s8 <-> fp32 */ - simple_reorder_t<f32, any, s8, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s8, any, f32, any, fmt_order::any, spec::reference>::pd_t::create, - /* u8 <-> fp32 */ - simple_reorder_t<f32, any, u8, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<u8, any, f32, any, fmt_order::any, spec::reference>::pd_t::create, - /* int{8,32} <-> int{8, 32} */ - simple_reorder_t<s8, any, s8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s8, any, s8, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<u8, any, u8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<u8, any, u8, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<s8, any, u8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s8, any, u8, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<u8, any, s8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<u8, any, s8, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<u8, any, s32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<u8, any, s32, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<s8, any, s32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s8, any, s32, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<s32, any, u8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s32, any, u8, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<s32, any, s8, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s32, any, s8, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<s32, any, s32, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s32, any, s32, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<s8, oihw, s8, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, oihw, s8, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, goihw, s8, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, goihw, s8, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, oihw, u8, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, oihw, u8, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, goihw, u8, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, goihw, u8, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, oihw, s8, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, oihw, s8, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, goihw, s8, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, goihw, s8, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, oihw, u8, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, oihw, u8, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, goihw, u8, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, goihw, u8, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, oihw, s32, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, oihw, s32, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, goihw, s32, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<u8, goihw, s32, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, oihw, s32, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, oihw, s32, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s8, goihw, s32, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s8, goihw, s32, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, oihw, u8, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, oihw, u8, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, goihw, u8, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, goihw, u8, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, oihw, s8, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, oihw, s8, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, goihw, s8, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, goihw, s8, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, oihw, s32, OIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, oihw, s32, OIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s32, goihw, s32, gOIhw4i16o4i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s32, goihw, s32, gOIhw4i16o4i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<u8, any, s8, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s8, any, u8, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<u8, any, s32, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s32, any, u8, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s8, any, s32, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s32, any, s8, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s32, any, s32, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<s8, any, s8, any, fmt_order::any, spec::reference>::pd_t::create, - simple_reorder_t<u8, any, u8, any, fmt_order::any, spec::reference>::pd_t::create, + + /* fp32: flat <-> blocked with tail */ + REG_SR_BIDIR(f32, any, f32, nChw8c), + REG_SR_BIDIR(f32, any, f32, nChw16c), + REG_SR_BIDIR(f32, any, f32, nCdhw16c), + REG_SR_BIDIR(f32, nChw8c, f32, nChw16c), + + REG_SR_BIDIR(f32, any, f32, Oihw16o), + REG_SR_BIDIR(f32, any, f32, Ohwi16o), + REG_SR_BIDIR(f32, any, f32, Oidhw16o), + REG_SR_BIDIR(f32, any, f32, Odhwi16o), + REG_SR_BIDIR(f32, any, f32, OIhw16o16i), + REG_SR_BIDIR(f32, any, f32, OIhw16i16o), + REG_SR_BIDIR(f32, any, f32, OIdhw16o16i), + REG_SR_BIDIR(f32, any, f32, OIdhw16i16o), + REG_SR_BIDIR(f32, any, f32, IOhw16o16i), + REG_SR_BIDIR(f32, any, f32, gOihw16o), + REG_SR_BIDIR(f32, any, f32, gOhwi16o), + REG_SR_BIDIR(f32, any, f32, gOidhw16o), + REG_SR_BIDIR(f32, any, f32, gOdhwi16o), + REG_SR_BIDIR(f32, any, f32, gOIhw16o16i), + REG_SR_BIDIR(f32, any, f32, gOIhw16i16o), + REG_SR_BIDIR(f32, any, f32, gOIdhw16o16i), + REG_SR_BIDIR(f32, any, f32, gOIdhw16i16o), + REG_SR_BIDIR(f32, any, f32, gIOhw16o16i), + + /* int: flat <-> blocked with tail */ + REG_SR_BIDIR(f32, nhwc, s32, nChw16c), + REG_SR_BIDIR(f32, nhwc, s8, nChw16c), + REG_SR_BIDIR(f32, nhwc, u8, nChw16c), + REG_SR_BIDIR(s32, nhwc, f32, nChw16c), + REG_SR_BIDIR(s32, nhwc, s32, nChw16c), + REG_SR_BIDIR(s32, nhwc, s8, nChw16c), + REG_SR_BIDIR(s32, nhwc, u8, nChw16c), + REG_SR_BIDIR(s8, nhwc, f32, nChw16c), + REG_SR_BIDIR(s8, nhwc, s32, nChw16c), + REG_SR_BIDIR(s8, nhwc, s8, nChw16c), + REG_SR_BIDIR(s8, nhwc, u8, nChw16c), + REG_SR_BIDIR(u8, nhwc, f32, nChw16c), + REG_SR_BIDIR(u8, nhwc, s32, nChw16c), + REG_SR_BIDIR(u8, nhwc, s8, nChw16c), + REG_SR_BIDIR(u8, nhwc, u8, nChw16c), + + REG_SR_BIDIR(f32, oihw, f32, OIhw4i16o4i), + REG_SR_BIDIR(f32, oihw, s8, OIhw4i16o4i), + REG_SR_BIDIR(s8, oihw, f32, OIhw4i16o4i), + REG_SR_BIDIR(s8, oihw, s8, OIhw4i16o4i), + REG_SR_BIDIR(f32, goihw, s8, gOIhw4i16o4i), + REG_SR_BIDIR(s8, goihw, f32, gOIhw4i16o4i), + REG_SR_BIDIR(f32, goihw, f32, gOIhw4i16o4i), + REG_SR_BIDIR(s8, goihw, s8, gOIhw4i16o4i), + /* s16 <-> s16 */ - simple_reorder_t<s16, any, s16, any, fmt_order::any, spec::direct_copy>::pd_t::create, - simple_reorder_t<s16, any, s16, any, fmt_order::any, spec::direct_copy_except_dim_0>::pd_t::create, - simple_reorder_t<s16, oihw, s16, OIhw8i16o2i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s16, oihw, s16, OIhw8i16o2i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s16, goihw, s16, gOIhw8i16o2i, fmt_order::keep>::pd_t::create, - simple_reorder_t<s16, goihw, s16, gOIhw8i16o2i, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s16, OIhw8i16o2i, s16, OIhw8o16i2o, fmt_order::keep>::pd_t::create, - simple_reorder_t<s16, OIhw8i16o2i, s16, OIhw8o16i2o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s16, gOIhw8i16o2i, s16, gOIhw8o16i2o, fmt_order::keep>::pd_t::create, - simple_reorder_t<s16, gOIhw8i16o2i, s16, gOIhw8o16i2o, fmt_order::reverse>::pd_t::create, - simple_reorder_t<s16, any, s16, any, fmt_order::any, spec::reference>::pd_t::create, + REG_SR_DIRECT_COPY(s16, s16), + REG_SR_BIDIR(s16, oihw, s16, OIhw8i16o2i), + REG_SR_BIDIR(s16, goihw, s16, gOIhw8i16o2i), + REG_SR_BIDIR(s16, OIhw8i16o2i, s16, OIhw8o16i2o), + REG_SR_BIDIR(s16, gOIhw8i16o2i, s16, gOIhw8o16i2o), + + /* WA to prevent fallback on reference implementations */ + REG_SR_DIRECT_COPY(u8, f32), + REG_SR_BIDIR(u8, nchw, f32, nChw8c), + REG_SR_BIDIR(u8, nchw, f32, nChw16c), + + /* reference: the last line of defence */ + REG_SR(f32, any, f32, any, fmt_order::any, spec::reference), + REG_SR(f32, any, s32, any, fmt_order::any, spec::reference), + REG_SR(f32, any, s16, any, fmt_order::any, spec::reference), + REG_SR(f32, any, s8, any, fmt_order::any, spec::reference), + REG_SR(f32, any, u8, any, fmt_order::any, spec::reference), + + REG_SR(s32, any, f32, any, fmt_order::any, spec::reference), + REG_SR(s32, any, s32, any, fmt_order::any, spec::reference), + REG_SR(s32, any, s16, any, fmt_order::any, spec::reference), + REG_SR(s32, any, s8, any, fmt_order::any, spec::reference), + REG_SR(s32, any, u8, any, fmt_order::any, spec::reference), + + REG_SR(s16, any, f32, any, fmt_order::any, spec::reference), + REG_SR(s16, any, s32, any, fmt_order::any, spec::reference), + REG_SR(s16, any, s16, any, fmt_order::any, spec::reference), + + REG_SR(s8, any, f32, any, fmt_order::any, spec::reference), + REG_SR(s8, any, s32, any, fmt_order::any, spec::reference), + REG_SR(s8, any, s8, any, fmt_order::any, spec::reference), + REG_SR(s8, any, u8, any, fmt_order::any, spec::reference), + + REG_SR(u8, any, f32, any, fmt_order::any, spec::reference), + REG_SR(u8, any, s32, any, fmt_order::any, spec::reference), + REG_SR(u8, any, u8, any, fmt_order::any, spec::reference), + REG_SR(u8, any, s8, any, fmt_order::any, spec::reference), + /* eol */ nullptr, }; |