added bench/LUT6

bench/LUT6.S

@@ -0,0 +1,101 @@
+#if MX_N == 4
+
+.global LUT6_rvv_vloxei8_m4
+LUT6_rvv_vloxei8_m4:
+	vsetvli t0, x0, e8, m4, ta, ma
+	li t0, 63
+	vmv.v.x v24, t0
+1:
+	vsetvli a3, a2, e8, m4, ta, ma
+	vle8.v v8, (a1)
+	vand.vv v8, v8, v24
+	vloxei8.v v8, (a0), v8
+	vse8.v v8, (a1)
+	sub a2, a2, a3
+	add a1, a1, a3
+	bnez a2, 1b
+	ret
+
+.global LUT6_rvv_vluxei8_m4
+LUT6_rvv_vluxei8_m4:
+	vsetvli t0, x0, e8, m4, ta, ma
+	li t0, 63
+	vmv.v.x v24, t0
+1:
+	vsetvli a3, a2, e8, m4, ta, ma
+	vle8.v v8, (a1)
+	vand.vv v8, v8, v24
+	vluxei8.v v8, (a0), v8
+	vse8.v v8, (a1)
+	sub a2, a2, a3
+	add a1, a1, a3
+	bnez a2, 1b
+	ret
+
+# a0 = lut, a1 = ptr, a2 = len
+.global LUT6_rvv_gather_m4
+LUT6_rvv_gather_m4:
+	li t0, 64
+	vsetvli zero, t0, e8, m4, ta, ma
+	vle8.v v0, (a0)
+
+	vsetvli t0, x0, e8, m4, ta, ma
+	li t0, 63
+	vmv.v.x v24, t0
+1:
+	vsetvli a0, a2, e8, m4, ta, ma
+	vle8.v v8, (a1)
+	vand.vv v8, v8, v24
+	vrgather.vv v16, v0, v8
+	vse8.v v16, (a1)
+	sub a2, a2, a0
+	add a1, a1, a0
+	bnez a2, 1b
+	ret
+
+.global LUT6_rvv_m1m2m4_gathers_m4
+LUT6_rvv_m1m2m4_gathers_m4:
+	li t0, 64
+	vsetvli zero, t0, e8, m4, ta, ma
+	vle8.v v0, (a0)
+
+	vsetvli t0, x0, e8, m4, ta, ma
+	li t0, 63
+	vmv.v.x v24, t0
+
+	csrr t0, vlenb
+	srl t0, t0, 4
+	sltiu t1, t0, 4
+	sltiu t0, t0, 2
+	j 0f
+1:
+	vsetvli t1, x0, e8, m1, ta, ma
+	vrgather.vv v16, v0, v8
+	vrgather.vv v17, v0, v9
+	vrgather.vv v18, v0, v10
+	vrgather.vv v19, v0, v11
+8:
+	vsetvli x0, a0, e8, m4, ta, ma
+	vse8.v v16, (a1)
+	sub a2, a2, a0
+	add a1, a1, a0
+	beqz a2, 9f
+0:
+	vsetvli a0, a2, e8, m4, ta, ma
+	vle8.v v8, (a1)
+	vand.vv v8, v8, v24
+	beqz t1, 1b
+	beqz t1, 2f
+	vrgather.vv v16, v0, v8
+	j 8b
+2:
+	vsetvli t1, x0, e8, m2, ta, ma
+	vrgather.vv v16, v0, v8
+	vrgather.vv v18, v0, v10
+	j 8b
+9:
+	ret
+
+#endif
+
+

bench/LUT6.c

@@ -0,0 +1,58 @@
+#include "bench.h"
+
+void
+LUT6_scalar(uint8_t lut[64], uint8_t *ptr, size_t n)
+{
+	for (; n--; ++ptr)
+		*ptr = lut[*ptr & 63], BENCH_CLOBBER();
+}
+
+void
+LUT6_scalar_autovec(uint8_t lut[64], uint8_t *ptr, size_t n)
+{
+	for (; n--; ++ptr)
+		*ptr = lut[*ptr & 63];
+}
+
+
+#define IMPLS(f) \
+	f(scalar) \
+	f(scalar_autovec) \
+	f(rvv_gather_m4) \
+	f(rvv_m1m2m4_gathers_m4) \
+	f(rvv_vluxei8_m4) \
+	f(rvv_vloxei8_m4) \
+
+typedef void Func(uint8_t lut[64], uint8_t *ptr, size_t n);
+
+#define DECLARE(f) extern Func LUT6_##f;
+IMPLS(DECLARE)
+
+#define EXTRACT(f) { #f, &LUT6_##f },
+Impl impls[] = { IMPLS(EXTRACT) };
+
+uint8_t *ptr;
+
+void init(void) { ptr = (uint8_t*)mem; }
+
+ux checksum(size_t n) {
+	ux sum = 0;
+	for (size_t i = 0; i < n; ++i)
+		sum = uhash(sum) + ptr[i];
+	return sum;
+}
+
+BENCH_BEG(base) {
+	static uint8_t lut[] =
+		"ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+		"abcdefghijklmnopqrstuvwxyz"
+		"0123456789"
+		"+/";
+	bench_memrand(ptr, n * sizeof *ptr);
+	TIME f(lut, ptr, n);
+} BENCH_END
+
+Bench benches[] = {
+	BENCH( impls, MAX_MEM, "LUT6", bench_base )
+}; BENCH_MAIN(benches)
+

bench/Makefile

@@ -2,7 +2,7 @@
 
 include ../config.mk
 
-EXECS=memcpy memset utf8_count strlen mergelines mandelbrot chacha20 poly1305 ascii_to_utf16 ascii_to_utf32 byteswap LUT4
+EXECS=memcpy memset utf8_count strlen mergelines mandelbrot chacha20 poly1305 ascii_to_utf16 ascii_to_utf32 byteswap LUT4 LUT6
 
 all: ${EXECS}
 

bench/bench.h

@@ -78,7 +78,7 @@ main(void)
 
 	size_t x;
 	randState.x ^= rv_cycles()*7;
-	randState.y += rv_cycles() ^ (uintptr_t)&x + 666*(uintptr_t)mem;
+	randState.y += rv_cycles() ^ ((uintptr_t)&x + 666*(uintptr_t)mem);
 
 	/* initialize memory */
 	bench_memrand(mem, MAX_MEM);