math/big: implement addVV in riscv64 assembly

This provides an assembly implementation of addVV for riscv64,
processing up to four words per loop, resulting in a significant
performance gain.

On a StarFive VisionFive 2:

               │   addvv.1    │               addvv.2               │
               │    sec/op    │   sec/op     vs base                │
AddVV/1-4         73.45n ± 0%   48.08n ± 0%  -34.54% (p=0.000 n=10)
AddVV/2-4         88.14n ± 0%   58.76n ± 0%  -33.33% (p=0.000 n=10)
AddVV/3-4        102.80n ± 0%   69.44n ± 0%  -32.45% (p=0.000 n=10)
AddVV/4-4        117.50n ± 0%   72.18n ± 0%  -38.57% (p=0.000 n=10)
AddVV/5-4        132.20n ± 0%   82.79n ± 0%  -37.38% (p=0.000 n=10)
AddVV/10-4        216.3n ± 0%   126.8n ± 0%  -41.35% (p=0.000 n=10)
AddVV/100-4      1659.0n ± 0%   885.2n ± 0%  -46.64% (p=0.000 n=10)
AddVV/1000-4     16.089µ ± 0%   8.400µ ± 0%  -47.79% (p=0.000 n=10)
AddVV/10000-4     245.3µ ± 0%   176.9µ ± 0%  -27.88% (p=0.000 n=10)
AddVV/100000-4    2.537m ± 0%   1.873m ± 0%  -26.17% (p=0.000 n=10)
geomean           1.435µ        904.5n       -36.99%

               │   addvv.1    │                addvv.2                │
               │     B/s      │      B/s       vs base                │
AddVV/1-4        830.9Mi ± 0%   1269.5Mi ± 0%  +52.78% (p=0.000 n=10)
AddVV/2-4        1.353Gi ± 0%    2.029Gi ± 0%  +50.00% (p=0.000 n=10)
AddVV/3-4        1.739Gi ± 0%    2.575Gi ± 0%  +48.09% (p=0.000 n=10)
AddVV/4-4        2.029Gi ± 0%    3.303Gi ± 0%  +62.82% (p=0.000 n=10)
AddVV/5-4        2.254Gi ± 0%    3.600Gi ± 0%  +59.69% (p=0.000 n=10)
AddVV/10-4       2.755Gi ± 0%    4.699Gi ± 0%  +70.54% (p=0.000 n=10)
AddVV/100-4      3.594Gi ± 0%    6.734Gi ± 0%  +87.37% (p=0.000 n=10)
AddVV/1000-4     3.705Gi ± 0%    7.096Gi ± 0%  +91.54% (p=0.000 n=10)
AddVV/10000-4    2.430Gi ± 0%    3.369Gi ± 0%  +38.65% (p=0.000 n=10)
AddVV/100000-4   2.350Gi ± 0%    3.183Gi ± 0%  +35.44% (p=0.000 n=10)
geomean          2.119Gi         3.364Gi       +58.71%

Change-Id: I727b3d9f8ab01eada7270046480b1430d56d0a96
Reviewed-on: https://go-review.googlesource.com/c/go/+/595395
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: M Zhuo <mengzhuo1203@gmail.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Than McIntosh <thanm@google.com>

diff --git a/src/math/big/arith_riscv64.s b/src/math/big/arith_riscv64.s
index bad32497b71c6739e7013199e2368355b8dc8686..67812dd646f7342db6571f100df6af0bbfa6eaa0 100644 (file)
--- a/src/math/big/arith_riscv64.s
+++ b/src/math/big/arith_riscv64.s
@@ -10,7 +10,86 @@
 // arithmetic operations on vectors implemented in arith.go.
 
 TEXT ·addVV(SB),NOSPLIT,$0
-       JMP ·addVV_g(SB)
+       MOV     x+24(FP), X5
+       MOV     y+48(FP), X6
+       MOV     z+0(FP), X7
+       MOV     z_len+8(FP), X30
+
+       MOV     $4, X28
+       MOV     $0, X29         // c = 0
+
+       BEQZ    X30, done
+       BLTU    X30, X28, loop1
+
+loop4:
+       MOV     0(X5), X8       // x[0]
+       MOV     0(X6), X9       // y[0]
+       MOV     8(X5), X11      // x[1]
+       MOV     8(X6), X12      // y[1]
+       MOV     16(X5), X14     // x[2]
+       MOV     16(X6), X15     // y[2]
+       MOV     24(X5), X17     // x[3]
+       MOV     24(X6), X18     // y[3]
+
+       ADD     X8, X9, X21     // z[0] = x[0] + y[0]
+       SLTU    X8, X21, X22
+       ADD     X21, X29, X10   // z[0] = x[0] + y[0] + c
+       SLTU    X21, X10, X23
+       ADD     X22, X23, X29   // next c
+
+       ADD     X11, X12, X24   // z[1] = x[1] + y[1]
+       SLTU    X11, X24, X25
+       ADD     X24, X29, X13   // z[1] = x[1] + y[1] + c
+       SLTU    X24, X13, X26
+       ADD     X25, X26, X29   // next c
+
+       ADD     X14, X15, X21   // z[2] = x[2] + y[2]
+       SLTU    X14, X21, X22
+       ADD     X21, X29, X16   // z[2] = x[2] + y[2] + c
+       SLTU    X21, X16, X23
+       ADD     X22, X23, X29   // next c
+
+       ADD     X17, X18, X21   // z[3] = x[3] + y[3]
+       SLTU    X17, X21, X22
+       ADD     X21, X29, X19   // z[3] = x[3] + y[3] + c
+       SLTU    X21, X19, X23
+       ADD     X22, X23, X29   // next c
+
+       MOV     X10, 0(X7)      // z[0]
+       MOV     X13, 8(X7)      // z[1]
+       MOV     X16, 16(X7)     // z[2]
+       MOV     X19, 24(X7)     // z[3]
+
+       ADD     $32, X5
+       ADD     $32, X6
+       ADD     $32, X7
+       SUB     $4, X30
+
+       BGEU    X30, X28, loop4
+       BEQZ    X30, done
+
+loop1:
+       MOV     0(X5), X10      // x
+       MOV     0(X6), X11      // y
+
+       ADD     X10, X11, X12   // z = x + y
+       SLTU    X10, X12, X14
+       ADD     X12, X29, X13   // z = x + y + c
+       SLTU    X12, X13, X15
+       ADD     X14, X15, X29   // next c
+
+       MOV     X13, 0(X7)      // z
+
+       ADD     $8, X5
+       ADD     $8, X6
+       ADD     $8, X7
+       SUB     $1, X30
+
+       BNEZ    X30, loop1
+
+done:
+       MOV     X29, c+72(FP)   // return c
+       RET
 
 TEXT ·subVV(SB),NOSPLIT,$0
        JMP ·subVV_g(SB)