simple reorder: disabled unused bf16 primitives

Author: antonvor

src/cpu/reorder/cpu_reorder_comp_s8_s8.cpp

@@ -127,7 +127,9 @@ const impl_list_map_t comp_s8_s8_impl_list_map {
         REG_REORDER_P(REG_SR(s8, ihwo, s8, OIhw16i16o4i, fmt_order_keep, spec_conv_req_comp))
         REG_REORDER_P(REG_SR(s8, oihw, s8, OIhw16i16o4i, fmt_order_keep, spec_conv_req_comp))
         REG_REORDER_P(REG_SR(s8, hwio, s8, OIhw16i16o4i, fmt_order_keep, spec_conv_req_comp))
+#endif
         REG_REORDER_P(REG_SR(s8, iohw, s8, OIhw4i16o4i, fmt_order_keep, spec_conv_req_comp))
+#ifdef ENABLE_UNUSED_PRIM
         REG_REORDER_P(REG_SR(s8, iohw, s8, OIhw2i8o4i, fmt_order_keep, spec_conv_req_comp))
         REG_REORDER_P(REG_SR(s8, iohw, s8, OIhw4o4i, fmt_order_keep, spec_conv_req_comp))
 #endif

src/cpu/reorder/cpu_reorder_regular_bf16.cpp

@@ -29,21 +29,26 @@ const impl_list_map_t regular_bf16_impl_list_map {
         REG_REORDER_P(DNNL_X64_ONLY(CPU_REORDER_INSTANCE(x64_jit_blk_reorder_t)))
         REG_REORDER_P(DNNL_X64_ONLY(CPU_REORDER_INSTANCE(x64_jit_uni_reorder_t)))
 
-// todo: [AV] disable unused simple reorder primitives
-        REG_REORDER_P(REG_SR_BIDIR(bf16, any, f32, nCw16c))
+#ifdef ENABLE_UNUSED_PRIM
+        REG_REORDER_P(REG_SR(bf16, any, f32, nCw16c, fmt_order_keep))
+#endif
+        REG_REORDER_P(REG_SR(bf16, any, f32, nCw16c, fmt_order_reverse))
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, f32, nChw16c))
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, f32, nCdhw16c))
-
+#ifdef ENABLE_UNUSED_PRIM
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, s8, nChw16c))
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, s8, nCdhw16c))
 
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, u8, nChw16c))
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, u8, nCdhw16c))
-
-        REG_REORDER_P(REG_SR_BIDIR(bf16, any, bf16, nCw16c))
+#endif
+        REG_REORDER_P(REG_SR(bf16, any, bf16, nCw16c, fmt_order_keep))
+#ifdef ENABLE_UNUSED_PRIM
+        REG_REORDER_P(REG_SR(bf16, any, bf16, nCw16c, fmt_order_reverse))
+#endif
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, bf16, nChw16c))
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, bf16, nCdhw16c))
-
+#ifdef ENABLE_UNUSED_PRIM
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, f32, OIdhw16o16i))
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, f32, OIdhw16i16o))
 
@@ -52,11 +57,13 @@ const impl_list_map_t regular_bf16_impl_list_map {
 
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, u8, OIdhw16o16i))
         REG_REORDER_P(REG_SR_BIDIR(bf16, any, u8, OIdhw16i16o))
-
+#endif
         REG_REORDER_P(REG_SR(bf16, any, bf16, any, fmt_order_any, spec_reference))
         REG_REORDER_P(REG_SR(bf16, any, f32, any, fmt_order_any, spec_reference))
+#ifdef ENABLE_UNUSED_PRIM
         REG_REORDER_P(REG_SR(bf16, any, s8, any, fmt_order_any, spec_reference))
         REG_REORDER_P(REG_SR(bf16, any, u8, any, fmt_order_any, spec_reference))
+#endif
 
         nullptr,
     }},

src/cpu/reorder/cpu_reorder_regular_f32_bf16.cpp

@@ -30,20 +30,32 @@ const impl_list_map_t regular_f32_bf16_impl_list_map {
         REG_REORDER_P(DNNL_X64_ONLY(CPU_REORDER_INSTANCE(x64_jit_blk_reorder_t)))
         REG_REORDER_P(DNNL_X64_ONLY(CPU_REORDER_INSTANCE(x64_jit_uni_reorder_t)))
 
-// todo: [AV] disable unused simple reorder primitives
-        REG_REORDER_P(REG_SR_BIDIR(f32, ncw, bf16, nCw16c))
-        REG_REORDER_P(REG_SR_BIDIR(f32, nchw, bf16, nChw16c))
+        REG_REORDER_P(REG_SR(f32, ncw, bf16, nCw16c, fmt_order_keep))
+#ifdef ENABLE_UNUSED_PRIM
+        REG_REORDER_P(REG_SR(f32, ncw, bf16, nCw16c, fmt_order_reverse))
+#endif
+        REG_REORDER_P(REG_SR(f32, nchw, bf16, nChw16c, fmt_order_keep))
+#ifdef ENABLE_UNUSED_PRIM
+        REG_REORDER_P(REG_SR(f32, nchw, bf16, nChw16c, fmt_order_reverse))
+#endif
         REG_REORDER_P(REG_SR_BIDIR(f32, any, bf16, nChw16c))
-        REG_REORDER_P(REG_SR_BIDIR(f32, any, bf16, nCdhw16c))
+        REG_REORDER_P(REG_SR(f32, any, bf16, nCdhw16c, fmt_order_keep))
+#ifdef ENABLE_UNUSED_PRIM
+        REG_REORDER_P(REG_SR(f32, any, bf16, nCdhw16c, fmt_order_reverse))
+#endif
 
         REG_REORDER_P(REG_SR(f32, oihw, bf16, OIhw8i16o2i, fmt_order_keep))
+#ifdef ENABLE_UNUSED_PRIM
         REG_REORDER_P(REG_SR(f32, goihw, bf16, gOIhw8i16o2i, fmt_order_keep))
+#endif
         REG_REORDER_P(REG_SR(f32, oihw, bf16, OIhw8o16i2o, fmt_order_keep))
+#ifdef ENABLE_UNUSED_PRIM
         REG_REORDER_P(REG_SR(f32, goihw, bf16, gOIhw8o16i2o, fmt_order_keep))
         REG_REORDER_P(REG_SR(f32, oihw, bf16, IOhw8o16i2o, fmt_order_keep))
         REG_REORDER_P(REG_SR(f32, goihw, bf16, gIOhw8o16i2o, fmt_order_keep))
         REG_REORDER_P(REG_SR(f32, oihw, bf16, OIhw16i16o, fmt_order_keep))
         REG_REORDER_P(REG_SR(f32, goihw, bf16, gOIhw16i16o, fmt_order_keep))
+#endif
 
         REG_REORDER_P(REG_SR(f32, any, bf16, any, fmt_order_any, spec_reference))
 

src/cpu/reorder/cpu_reorder_regular_u8.cpp

@@ -57,10 +57,8 @@ const impl_list_map_t regular_u8_impl_list_map {
         REG_REORDER_P(REG_SR_BIDIR(u8, any, s32, nChw16c))
         REG_REORDER_P(REG_SR_BIDIR(u8, any, bf16, nChw16c))
         REG_REORDER_P(REG_SR_BIDIR(u8, any, s8, nChw16c))
-
-        REG_REORDER_P(REG_SR(u8, any, u8, nChw16c, fmt_order_keep))
 #endif
-        REG_REORDER_P(REG_SR(u8, any, u8, nChw16c, fmt_order_reverse))
+        REG_REORDER_P(REG_SR_BIDIR(u8, any, u8, nChw16c))
         REG_REORDER_P(REG_SR(u8, any, f32, nChw8c, fmt_order_keep))
 #ifdef ENABLE_UNUSED_PRIM
         REG_REORDER_P(REG_SR(u8, any, f32, nChw8c, fmt_order_reverse))