Use vnclip+max instead of min+max+vncvt

imaginationtech · Sep 25, 2024 · 763c584 · 763c584
1 parent bb3d9e9
commit 763c584
Show file tree

Hide file tree

Showing 11 changed files with 22 additions and 32 deletions.
diff --git a/src/configs/binary-elementwise-config.c b/src/configs/binary-elementwise-config.c
@@ -965,7 +965,7 @@ static void init_qs8_vadd_config(void) {
     qs8_vadd_config.op_ukernel = (xnn_vbinary_ukernel_fn) xnn_qs8_vadd_minmax_ukernel__rvv_u2v;
     qs8_vadd_config.opc_ukernel = (xnn_vbinary_ukernel_fn) xnn_qs8_vaddc_minmax_ukernel__rvv_u2v;
     qs8_vadd_config.ropc_ukernel = (xnn_vbinary_ukernel_fn) xnn_qs8_vaddc_minmax_ukernel__rvv_u2v;
-    qs8_vadd_config.init = xnn_init_qs8_add_minmax_scalar_params;
+    qs8_vadd_config.init = (xnn_init_binary_params_fn) xnn_init_qs8_add_minmax_scalar_params;
     qs8_vadd_config.element_tile = hardware_config->vlenb;
   #else
     qs8_vadd_config.op_ukernel = (xnn_vbinary_ukernel_fn) xnn_qs8_vadd_minmax_ukernel__scalar_u4;
@@ -1110,7 +1110,7 @@ static void init_qu8_vadd_config(void) {
     qu8_vadd_config.op_ukernel = (xnn_vbinary_ukernel_fn) xnn_qu8_vadd_minmax_ukernel__rvv_u2v;
     qu8_vadd_config.opc_ukernel = (xnn_vbinary_ukernel_fn) xnn_qu8_vaddc_minmax_ukernel__rvv_u2v;
     qu8_vadd_config.ropc_ukernel = (xnn_vbinary_ukernel_fn) xnn_qu8_vaddc_minmax_ukernel__rvv_u2v;
-    qu8_vadd_config.init = xnn_init_qu8_add_minmax_scalar_params;
+    qu8_vadd_config.init = (xnn_init_binary_params_fn) xnn_init_qu8_add_minmax_scalar_params;
     qu8_vadd_config.element_tile = hardware_config->vlenb;
   #else
     qu8_vadd_config.op_ukernel = (xnn_vbinary_ukernel_fn) xnn_qu8_vadd_minmax_ukernel__scalar_u4;

diff --git a/src/qs8-vadd/rvv.c.in b/src/qs8-vadd/rvv.c.in
@@ -57,9 +57,8 @@ void xnn_${DATATYPE.lower()}_vadd_minmax_ukernel__rvv_u${LMUL}v(
     acc_i32v = __riscv_vadd_vv_i32m${LMUL*4}(acc_i32v, b_i32v, n);
     vint32m${LMUL*4}_t out_i32v = __riscv_vsra_vx_i32m${LMUL*4}(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m${LMUL*4}(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m${LMUL*4}(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m${LMUL*4}(out_i32v, output_max, n);
-    vint16m${LMUL*2}_t out_i16v = __riscv_vncvt_x_x_w_i16m${LMUL*2}(out_i32v, n);
+    vint16m${LMUL*2}_t out_i16v = __riscv_vnclip_wx_i16m${LMUL*2}(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m${LMUL*2}(out_i16v, output_min, n);
     $if DATATYPE == "QS8":
       vint8m${LMUL}_t out_i8v = __riscv_vncvt_x_x_w_i8m${LMUL}(out_i16v, n);
       __riscv_vse8_v_i8m${LMUL}(output, out_i8v, n); output += n;

diff --git a/src/qs8-vaddc/rvv.c.in b/src/qs8-vaddc/rvv.c.in
@@ -48,9 +48,8 @@ void xnn_${DATATYPE.lower()}_vaddc_minmax_ukernel__rvv_u${LMUL}v(
     vint32m${LMUL*4}_t acc_i32v = __riscv_vadd_vx_i32m${LMUL*4}(a_i32v, bias, n);
     vint32m${LMUL*4}_t out_i32v = __riscv_vsra_vx_i32m${LMUL*4}(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m${LMUL*4}(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m${LMUL*4}(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m${LMUL*4}(out_i32v, output_max, n);
-    vint16m${LMUL*2}_t out_i16v = __riscv_vncvt_x_x_w_i16m${LMUL*2}(out_i32v, n);
+    vint16m${LMUL*2}_t out_i16v = __riscv_vnclip_wx_i16m${LMUL*2}(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m${LMUL*2}(out_i16v, output_min, n);
     $if DATATYPE == "QS8":
       vint8m${LMUL}_t out_i8v = __riscv_vncvt_x_x_w_i8m${LMUL}(out_i16v, n);
       __riscv_vse8_v_i8m${LMUL}(output, out_i8v, n); output += n;

diff --git a/src/qs8-vmul/gen/qs8-vadd-minmax-rvv-u1v.c b/src/qs8-vmul/gen/qs8-vadd-minmax-rvv-u1v.c
@@ -50,9 +50,8 @@ void xnn_qs8_vadd_minmax_ukernel__rvv_u1v(
     acc_i32v = __riscv_vadd_vv_i32m4(acc_i32v, b_i32v, n);
     vint32m4_t out_i32v = __riscv_vsra_vx_i32m4(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m4(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m4(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m4(out_i32v, output_max, n);
-    vint16m2_t out_i16v = __riscv_vncvt_x_x_w_i16m2(out_i32v, n);
+    vint16m2_t out_i16v = __riscv_vnclip_wx_i16m2(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m2(out_i16v, output_min, n);
     vint8m1_t out_i8v = __riscv_vncvt_x_x_w_i8m1(out_i16v, n);
     __riscv_vse8_v_i8m1(output, out_i8v, n); output += n;
   } while (batch != 0);

diff --git a/src/qs8-vmul/gen/qs8-vadd-minmax-rvv-u2v.c b/src/qs8-vmul/gen/qs8-vadd-minmax-rvv-u2v.c
@@ -50,9 +50,8 @@ void xnn_qs8_vadd_minmax_ukernel__rvv_u2v(
     acc_i32v = __riscv_vadd_vv_i32m8(acc_i32v, b_i32v, n);
     vint32m8_t out_i32v = __riscv_vsra_vx_i32m8(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m8(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m8(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m8(out_i32v, output_max, n);
-    vint16m4_t out_i16v = __riscv_vncvt_x_x_w_i16m4(out_i32v, n);
+    vint16m4_t out_i16v = __riscv_vnclip_wx_i16m4(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m4(out_i16v, output_min, n);
     vint8m2_t out_i8v = __riscv_vncvt_x_x_w_i8m2(out_i16v, n);
     __riscv_vse8_v_i8m2(output, out_i8v, n); output += n;
   } while (batch != 0);

diff --git a/src/qs8-vmulc/gen/qs8-vaddc-minmax-rvv-u1v.c b/src/qs8-vmulc/gen/qs8-vaddc-minmax-rvv-u1v.c
@@ -44,9 +44,8 @@ void xnn_qs8_vaddc_minmax_ukernel__rvv_u1v(
     vint32m4_t acc_i32v = __riscv_vadd_vx_i32m4(a_i32v, bias, n);
     vint32m4_t out_i32v = __riscv_vsra_vx_i32m4(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m4(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m4(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m4(out_i32v, output_max, n);
-    vint16m2_t out_i16v = __riscv_vncvt_x_x_w_i16m2(out_i32v, n);
+    vint16m2_t out_i16v = __riscv_vnclip_wx_i16m2(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m2(out_i16v, output_min, n);
     vint8m1_t out_i8v = __riscv_vncvt_x_x_w_i8m1(out_i16v, n);
     __riscv_vse8_v_i8m1(output, out_i8v, n); output += n;
   } while (batch != 0);

diff --git a/src/qs8-vmulc/gen/qs8-vaddc-minmax-rvv-u2v.c b/src/qs8-vmulc/gen/qs8-vaddc-minmax-rvv-u2v.c
@@ -44,9 +44,8 @@ void xnn_qs8_vaddc_minmax_ukernel__rvv_u2v(
     vint32m8_t acc_i32v = __riscv_vadd_vx_i32m8(a_i32v, bias, n);
     vint32m8_t out_i32v = __riscv_vsra_vx_i32m8(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m8(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m8(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m8(out_i32v, output_max, n);
-    vint16m4_t out_i16v = __riscv_vncvt_x_x_w_i16m4(out_i32v, n);
+    vint16m4_t out_i16v = __riscv_vnclip_wx_i16m4(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m4(out_i16v, output_min, n);
     vint8m2_t out_i8v = __riscv_vncvt_x_x_w_i8m2(out_i16v, n);
     __riscv_vse8_v_i8m2(output, out_i8v, n); output += n;
   } while (batch != 0);

diff --git a/src/qu8-vmul/gen/qu8-vadd-minmax-rvv-u1v.c b/src/qu8-vmul/gen/qu8-vadd-minmax-rvv-u1v.c
@@ -52,9 +52,8 @@ void xnn_qu8_vadd_minmax_ukernel__rvv_u1v(
     acc_i32v = __riscv_vadd_vv_i32m4(acc_i32v, b_i32v, n);
     vint32m4_t out_i32v = __riscv_vsra_vx_i32m4(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m4(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m4(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m4(out_i32v, output_max, n);
-    vint16m2_t out_i16v = __riscv_vncvt_x_x_w_i16m2(out_i32v, n);
+    vint16m2_t out_i16v = __riscv_vnclip_wx_i16m2(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m2(out_i16v, output_min, n);
     a_u16v = __riscv_vreinterpret_v_i16m2_u16m2(out_i16v);
     vuint8m1_t out_u8v = __riscv_vncvt_x_x_w_u8m1(a_u16v, n);
     __riscv_vse8_v_u8m1(output, out_u8v, n); output += n;

diff --git a/src/qu8-vmul/gen/qu8-vadd-minmax-rvv-u2v.c b/src/qu8-vmul/gen/qu8-vadd-minmax-rvv-u2v.c
@@ -52,9 +52,8 @@ void xnn_qu8_vadd_minmax_ukernel__rvv_u2v(
     acc_i32v = __riscv_vadd_vv_i32m8(acc_i32v, b_i32v, n);
     vint32m8_t out_i32v = __riscv_vsra_vx_i32m8(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m8(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m8(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m8(out_i32v, output_max, n);
-    vint16m4_t out_i16v = __riscv_vncvt_x_x_w_i16m4(out_i32v, n);
+    vint16m4_t out_i16v = __riscv_vnclip_wx_i16m4(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m4(out_i16v, output_min, n);
     a_u16v = __riscv_vreinterpret_v_i16m4_u16m4(out_i16v);
     vuint8m2_t out_u8v = __riscv_vncvt_x_x_w_u8m2(a_u16v, n);
     __riscv_vse8_v_u8m2(output, out_u8v, n); output += n;

diff --git a/src/qu8-vmulc/gen/qu8-vaddc-minmax-rvv-u1v.c b/src/qu8-vmulc/gen/qu8-vaddc-minmax-rvv-u1v.c
@@ -45,9 +45,8 @@ void xnn_qu8_vaddc_minmax_ukernel__rvv_u1v(
     vint32m4_t acc_i32v = __riscv_vadd_vx_i32m4(a_i32v, bias, n);
     vint32m4_t out_i32v = __riscv_vsra_vx_i32m4(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m4(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m4(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m4(out_i32v, output_max, n);
-    vint16m2_t out_i16v = __riscv_vncvt_x_x_w_i16m2(out_i32v, n);
+    vint16m2_t out_i16v = __riscv_vnclip_wx_i16m2(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m2(out_i16v, output_min, n);
     a_u16v = __riscv_vreinterpret_v_i16m2_u16m2(out_i16v);
     vuint8m1_t out_u8v = __riscv_vncvt_x_x_w_u8m1(a_u16v, n);
     __riscv_vse8_v_u8m1(output, out_u8v, n); output += n;

diff --git a/src/qu8-vmulc/gen/qu8-vaddc-minmax-rvv-u2v.c b/src/qu8-vmulc/gen/qu8-vaddc-minmax-rvv-u2v.c
@@ -45,9 +45,8 @@ void xnn_qu8_vaddc_minmax_ukernel__rvv_u2v(
     vint32m8_t acc_i32v = __riscv_vadd_vx_i32m8(a_i32v, bias, n);
     vint32m8_t out_i32v = __riscv_vsra_vx_i32m8(acc_i32v, shift, n);
     out_i32v = __riscv_vadd_vx_i32m8(out_i32v, output_zero_point, n);
-    out_i32v = __riscv_vmax_vx_i32m8(out_i32v, output_min, n);
-    out_i32v = __riscv_vmin_vx_i32m8(out_i32v, output_max, n);
-    vint16m4_t out_i16v = __riscv_vncvt_x_x_w_i16m4(out_i32v, n);
+    vint16m4_t out_i16v = __riscv_vnclip_wx_i16m4(out_i32v, output_max, __RISCV_VXRM_RNE, n);
+    out_i16v = __riscv_vmax_vx_i16m4(out_i16v, output_min, n);
     a_u16v = __riscv_vreinterpret_v_i16m4_u16m4(out_i16v);
     vuint8m2_t out_u8v = __riscv_vncvt_x_x_w_u8m2(a_u16v, n);
     __riscv_vse8_v_u8m2(output, out_u8v, n); output += n;