@@ -32,11 +32,15 @@ static struct xnn_binary_elementwise_config f32_vprelu_config = {0};
3232static struct xnn_binary_elementwise_config f32_vsub_config = {0 };
3333static struct xnn_binary_elementwise_config f32_vsqrdiff_config = {0 };
3434
35+
3536static struct xnn_binary_elementwise_config qs8_vadd_config = {0 };
3637static struct xnn_binary_elementwise_config qs8_vmul_config = {0 };
38+ static struct xnn_binary_elementwise_config qs8_vprelu_config = {0 };
39+
3740
3841static struct xnn_binary_elementwise_config qu8_vadd_config = {0 };
3942static struct xnn_binary_elementwise_config qu8_vmul_config = {0 };
43+ static struct xnn_binary_elementwise_config qu8_vprelu_config = {0 };
4044
4145XNN_INIT_ONCE_GUARD (f16_vadd );
4246XNN_INIT_ONCE_GUARD (f16_vdiv );
@@ -57,8 +61,10 @@ XNN_INIT_ONCE_GUARD(f32_vsub);
5761XNN_INIT_ONCE_GUARD (f32_vsqrdiff );
5862XNN_INIT_ONCE_GUARD (qs8_vadd );
5963XNN_INIT_ONCE_GUARD (qs8_vmul );
64+ XNN_INIT_ONCE_GUARD (qs8_vprelu );
6065XNN_INIT_ONCE_GUARD (qu8_vadd );
6166XNN_INIT_ONCE_GUARD (qu8_vmul );
67+ XNN_INIT_ONCE_GUARD (qu8_vprelu );
6268
6369
6470static void init_f16_vadd_config (void ) {
@@ -1105,6 +1111,30 @@ static void init_qs8_vmul_config(void) {
11051111 #endif
11061112}
11071113
1114+ static void init_qs8_vprelu_config (void ) {
1115+ #if XNN_ARCH_X86 || XNN_ARCH_X86_64
1116+ const struct xnn_hardware_config * hardware_config = xnn_init_hardware_config ();
1117+ assert (hardware_config != NULL );
1118+ if (hardware_config -> use_x86_avx2 )
1119+ {
1120+ qs8_vprelu_config .op_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qs8_vprelu_ukernel__avx2_u16 ;
1121+ qs8_vprelu_config .opc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qs8_vpreluc_ukernel__avx2_u16 ;
1122+ qs8_vprelu_config .ropc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qs8_vpreluc_ukernel__avx2_u16 ;
1123+ qs8_vprelu_config .init = (xnn_init_binary_params_fn ) xnn_init_qs8_vprelu_scalar_params ;
1124+ qs8_vprelu_config .element_tile = 16 ;
1125+ }
1126+ else
1127+ {
1128+ qs8_vprelu_config .op_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qs8_vprelu_ukernel__scalar_u8 ;
1129+ qs8_vprelu_config .opc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qs8_vpreluc_ukernel__scalar_u8 ;
1130+ qs8_vprelu_config .ropc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qs8_vpreluc_ukernel__scalar_u8 ;
1131+ qs8_vprelu_config .init = (xnn_init_binary_params_fn ) xnn_init_qs8_vprelu_scalar_params ;
1132+ qs8_vprelu_config .element_tile = 8 ;
1133+ }
1134+
1135+ #endif
1136+ }
1137+
11081138static void init_qu8_vadd_config (void ) {
11091139 #if XNN_ARCH_ARM
11101140 const struct xnn_hardware_config * hardware_config = xnn_init_hardware_config ();
@@ -1246,6 +1276,30 @@ static void init_qu8_vmul_config(void) {
12461276 #endif
12471277}
12481278
1279+ static void init_qu8_vprelu_config (void ) {
1280+ #if XNN_ARCH_X86 || XNN_ARCH_X86_64
1281+ const struct xnn_hardware_config * hardware_config = xnn_init_hardware_config ();
1282+ assert (hardware_config != NULL );
1283+ if (hardware_config -> use_x86_avx2 )
1284+ {
1285+ qu8_vprelu_config .op_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qu8_vprelu_ukernel__avx2_u16 ;
1286+ qu8_vprelu_config .opc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qu8_vpreluc_ukernel__avx2_u16 ;
1287+ qu8_vprelu_config .ropc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qu8_vpreluc_ukernel__avx2_u16 ;
1288+ qu8_vprelu_config .init = (xnn_init_binary_params_fn ) xnn_init_qu8_vprelu_scalar_params ;
1289+ qu8_vprelu_config .element_tile = 16 ;
1290+ }
1291+ else
1292+ {
1293+ qu8_vprelu_config .op_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qu8_vprelu_ukernel__scalar_u8 ;
1294+ qu8_vprelu_config .opc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qu8_vpreluc_ukernel__scalar_u8 ;
1295+ qu8_vprelu_config .ropc_ukernel = (xnn_vbinary_ukernel_fn ) xnn_qu8_vpreluc_ukernel__scalar_u8 ;
1296+ qu8_vprelu_config .init = (xnn_init_binary_params_fn ) xnn_init_qu8_vprelu_scalar_params ;
1297+ qu8_vprelu_config .element_tile = 8 ;
1298+ }
1299+
1300+ #endif
1301+ }
1302+
12491303const struct xnn_binary_elementwise_config * xnn_init_f16_vadd_config () {
12501304 const struct xnn_hardware_config * hardware_config = xnn_init_hardware_config ();
12511305 if (hardware_config == NULL || !xnn_is_f16_compatible_config (hardware_config )) {
@@ -1417,6 +1471,15 @@ const struct xnn_binary_elementwise_config* xnn_init_qs8_vmul_config() {
14171471 return & qs8_vmul_config ;
14181472}
14191473
1474+ const struct xnn_binary_elementwise_config * xnn_init_qs8_vprelu_config () {
1475+ const struct xnn_hardware_config * hardware_config = xnn_init_hardware_config ();
1476+ if (hardware_config == NULL ) {
1477+ return NULL ;
1478+ }
1479+ XNN_INIT_ONCE (qs8_vprelu );
1480+ return & qs8_vprelu_config ;
1481+ }
1482+
14201483const struct xnn_binary_elementwise_config * xnn_init_qu8_vadd_config () {
14211484 const struct xnn_hardware_config * hardware_config = xnn_init_hardware_config ();
14221485 if (hardware_config == NULL ) {
@@ -1434,3 +1497,12 @@ const struct xnn_binary_elementwise_config* xnn_init_qu8_vmul_config() {
14341497 XNN_INIT_ONCE (qu8_vmul );
14351498 return & qu8_vmul_config ;
14361499}
1500+
1501+ const struct xnn_binary_elementwise_config * xnn_init_qu8_vprelu_config () {
1502+ const struct xnn_hardware_config * hardware_config = xnn_init_hardware_config ();
1503+ if (hardware_config == NULL ) {
1504+ return NULL ;
1505+ }
1506+ XNN_INIT_ONCE (qu8_vprelu );
1507+ return & qu8_vprelu_config ;
1508+ }
0 commit comments