https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118019
--- Comment #4 from Li Pan <pan2.li at intel dot com> ---
(In reply to Li Pan from comment #3)
> 1 │ #include <stdint.h>
> 2 │
> 3 │ #define I_P1 16
> 4 │ #define I_P2 1344
> 5 │
> 6 │ #define HADAMARD4(d0, d1, d2, d3, s0, s1, s2, s3) {\
> 7 │ int t0 = s0 + s1;\
> 8 │ int t1 = s0 - s1;\
> 9 │ int t2 = s2 + s3;\
> 10 │ int t3 = s2 - s3;\
> 11 │ d0 = t0 + t2;\
> 12 │ d2 = t0 - t2;\
> 13 │ d1 = t1 + t3;\
> 14 │ d3 = t1 - t3;\
> 15 │ }
> 16 │
> 17 │ uint32_t abs2( uint32_t a )
> 18 │ {
> 19 │ uint32_t s = ((a >> 15) & 0x10001) * 0xffff;
> 20 │ return (a+s)^s;
> 21 │ }
> 22 │
> 23 │ int x264_pixel_satd_8x4 (uint8_t * pix1, int i_pix1, uint8_t *
> pix2, int i_pix2)
> 24 │ {
> 25 │ uint32_t tmp[4][4];
> 26 │ uint32_t a0, a1, a2, a3;
> 27 │ int sum = 0;
> 28 │
> 29 │ for (int i = 0; i < 4; i++, pix1 += i_pix1, pix2 += i_pix2) {
> 30 │ a0 = (pix1[0] - pix2[0]) + ((pix1[4] - pix2[4]) << 16);
> 31 │ a1 = (pix1[1] - pix2[1]) + ((pix1[5] - pix2[5]) << 16);
> 32 │ a2 = (pix1[2] - pix2[2]) + ((pix1[6] - pix2[6]) << 16);
> 33 │ a3 = (pix1[3] - pix2[3]) + ((pix1[7] - pix2[7]) << 16);
> 34 │
> 35 │ HADAMARD4 (tmp[i][0], tmp[i][1], tmp[i][2], tmp[i][3],
> a0,a1,a2,a3);
> 36 │ }
> 37 │
> 38 │ for( int i = 0; i < 4; i++ ) {
> 39 │ HADAMARD4 (a0, a1, a2, a3, tmp[0][i], tmp[1][i], tmp[2][i],
> tmp[3][i]);
> 40 │ sum += abs2 (a0) + abs2 (a1) + abs2 (a2) + abs2 (a3);
> 41 │ }
> 42 │
> 43 │ return (((uint16_t)sum) + ((uint32_t)sum >> 16)) >> 1;
> 44 │ }
> 45 │
> 46 │ __attribute__((aligned(16))) uint64_t g_pix1[I_P2 * 5] = { 111,
> 112, 113, 114, 115, 116, 117, 118 };
> 47 │ __attribute__((aligned(16))) uint64_t g_pix2[I_P2 * 5] = { 211,
> 212, 213, 214, 215, 216, 217, 218 };
> 48 │ int result;
> 49 │
> 50 │ int main() {
> 51 │ result = x264_pixel_satd_8x4((uint8_t*)g_pix1, I_P1,
> (uint8_t*)g_pix2, I_P2);
> 52 │ return 0;
> 53 │ }
>
> ../__RV64/bin/riscv64-unknown-elf-gcc satd.c -march=rv64gcv_zvl512b
> -mrvv-vector-bits=zvl -O3 -g0 -mrvv-max-lmul=m2 -mvector-strict-align -o
> test.salign.elf
>
> QEMU_CPU=rv64,vlen=512,rvv_ta_all_1s=true,rvv_ma_all_1s=true,v=true,
> vext_spec=v1.0 ~/bin/qemu/bin/qemu-riscv64 -plugin
> /home/pli/repos/qemu/build/tests/tcg/plugins/libinsn.so -d plugin
> test.salign.elf
>
> total insns: 917
>
> ../__RV64/bin/riscv64-unknown-elf-gcc satd.c -march=rv64gcv_zvl512b
> -mrvv-vector-bits=zvl -O3 -g0 -mrvv-max-lmul=m2 -mno-vector-strict-align -o
> test.no-salign.elf
>
> QEMU_CPU=rv64,vlen=512,rvv_ta_all_1s=true,rvv_ma_all_1s=true,v=true,
> vext_spec=v1.0 ~/bin/qemu/bin/qemu-riscv64 -plugin
> /home/pli/repos/qemu/build/tests/tcg/plugins/libinsn.so -d plugin
> test.no-salign.elf
>
> total insns: 755
Then turn to gcc-14.
~/bin/gcc-14/bin/riscv64-unknown-elf-gcc satd.c -march=rv64gcv_zvl512b
-mrvv-vector-bits=zvl -O3 -g0 -mrvv-max-lmul=m2 -static -o satd.gcc.14.elf
QEMU_CPU=rv64,vlen=512,rvv_ta_all_1s=true,rvv_ma_all_1s=true,v=true,vext_spec=v1.0
~/bin/qemu/bin/qemu-riscv64 -plugin
/home/pli/repos/qemu/build/tests/tcg/plugins/libinsn.so -d plugin
satd.gcc.14.elf
total insns: 829
So for the option "-march=rv64gcv_zvl512b -mrvv-vector-bits=zvl -O3 -g0
-mrvv-max-lmul=m2", upstream insn count is worse than gcc-14.
