Enable S8/U8_VCLAMP microkernels for RISC-V

PiperOrigin-RevId: 688619523
google · Oct 22, 2024 · 31205d0 · 31205d0
1 parent e9975bc
commit 31205d0
Show file tree

Hide file tree

Showing 3 changed files with 11 additions and 4 deletions.
diff --git a/cmake/gen/rvv_microkernels.cmake b/cmake/gen/rvv_microkernels.cmake
@@ -62,6 +62,8 @@ SET(PROD_RVV_MICROKERNEL_SRCS
   src/qu8-vlrelu/gen/qu8-vlrelu-rvv-u2v.c
   src/qu8-vmul/gen/qu8-vmul-minmax-f32-rvv-u2v.c
   src/qu8-vmulc/gen/qu8-vmulc-minmax-f32-rvv-u2v.c
+  src/s8-vclamp/gen/s8-vclamp-rvv-u4v.c
+  src/u8-vclamp/gen/u8-vclamp-rvv-u4v.c
   src/x32-packw/gen/x32-packw-x4v-gemm-goi-rvv-u8.c
   src/x32-transposec/gen/x32-transposec-4x4-rvv.c
   src/x32-transposec/gen/x32-transposec-8x8-rvv.c
@@ -192,11 +194,9 @@ SET(NON_PROD_RVV_MICROKERNEL_SRCS
   src/qu8-vmulc/gen/qu8-vmulc-minmax-f32-rvv-u1v.c
   src/s8-vclamp/gen/s8-vclamp-rvv-u1v.c
   src/s8-vclamp/gen/s8-vclamp-rvv-u2v.c
-  src/s8-vclamp/gen/s8-vclamp-rvv-u4v.c
   src/s8-vclamp/gen/s8-vclamp-rvv-u8v.c
   src/u8-vclamp/gen/u8-vclamp-rvv-u1v.c
   src/u8-vclamp/gen/u8-vclamp-rvv-u2v.c
-  src/u8-vclamp/gen/u8-vclamp-rvv-u4v.c
   src/u8-vclamp/gen/u8-vclamp-rvv-u8v.c
   src/x32-packw/gen/x32-packw-x1v-gemm-goi-rvv-u2.c
   src/x32-packw/gen/x32-packw-x1v-gemm-goi-rvv-u4.c

diff --git a/gen/rvv_microkernels.bzl b/gen/rvv_microkernels.bzl
@@ -58,6 +58,8 @@ PROD_RVV_MICROKERNEL_SRCS = [
     "src/qu8-vlrelu/gen/qu8-vlrelu-rvv-u2v.c",
     "src/qu8-vmul/gen/qu8-vmul-minmax-f32-rvv-u2v.c",
     "src/qu8-vmulc/gen/qu8-vmulc-minmax-f32-rvv-u2v.c",
+    "src/s8-vclamp/gen/s8-vclamp-rvv-u4v.c",
+    "src/u8-vclamp/gen/u8-vclamp-rvv-u4v.c",
     "src/x32-packw/gen/x32-packw-x4v-gemm-goi-rvv-u8.c",
     "src/x32-transposec/gen/x32-transposec-4x4-rvv.c",
     "src/x32-transposec/gen/x32-transposec-8x8-rvv.c",
@@ -189,11 +191,9 @@ NON_PROD_RVV_MICROKERNEL_SRCS = [
     "src/qu8-vmulc/gen/qu8-vmulc-minmax-f32-rvv-u1v.c",
     "src/s8-vclamp/gen/s8-vclamp-rvv-u1v.c",
     "src/s8-vclamp/gen/s8-vclamp-rvv-u2v.c",
-    "src/s8-vclamp/gen/s8-vclamp-rvv-u4v.c",
     "src/s8-vclamp/gen/s8-vclamp-rvv-u8v.c",
     "src/u8-vclamp/gen/u8-vclamp-rvv-u1v.c",
     "src/u8-vclamp/gen/u8-vclamp-rvv-u2v.c",
-    "src/u8-vclamp/gen/u8-vclamp-rvv-u4v.c",
     "src/u8-vclamp/gen/u8-vclamp-rvv-u8v.c",
     "src/x32-packw/gen/x32-packw-x1v-gemm-goi-rvv-u2.c",
     "src/x32-packw/gen/x32-packw-x1v-gemm-goi-rvv-u4.c",

diff --git a/src/configs/unary-elementwise-config.c b/src/configs/unary-elementwise-config.c
@@ -1979,6 +1979,10 @@ static void init_s8_clamp_config(void) {
   #elif XNN_ARCH_ARM64
     s8_clamp_config.ukernel = (xnn_vunary_ukernel_fn) xnn_s8_vclamp_ukernel__neon_u64;
     s8_clamp_config.init.s8_minmax = xnn_init_s8_minmax_scalar_params;
+
+  #elif XNN_ARCH_RISCV && XNN_ENABLE_RISCV_VECTOR
+    s8_clamp_config.ukernel = (xnn_vunary_ukernel_fn) xnn_s8_vclamp_ukernel__rvv_u4v;
+    s8_clamp_config.init.s8_minmax = xnn_init_s8_minmax_scalar_params;
   #elif XNN_ARCH_X86 || XNN_ARCH_X86_64
     const struct xnn_hardware_config* hardware_config = xnn_init_hardware_config();
     assert(hardware_config != NULL);
@@ -2021,6 +2025,9 @@ static void init_u8_clamp_config(void) {
   #elif XNN_ARCH_ARM64
     u8_clamp_config.ukernel = (xnn_vunary_ukernel_fn) xnn_u8_vclamp_ukernel__neon_u64;
     u8_clamp_config.init.u8_minmax = xnn_init_u8_minmax_scalar_params;
+  #elif XNN_ARCH_RISCV && XNN_ENABLE_RISCV_VECTOR
+    u8_clamp_config.ukernel = (xnn_vunary_ukernel_fn) xnn_u8_vclamp_ukernel__rvv_u4v;
+    u8_clamp_config.init.u8_minmax = xnn_init_u8_minmax_scalar_params;
   #elif XNN_ARCH_X86 || XNN_ARCH_X86_64
     const struct xnn_hardware_config* hardware_config = xnn_init_hardware_config();
     assert(hardware_config != NULL);