#include "macros.inc" .align 2 .global PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_top .global _PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_top #ifndef __clang__ .type PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_top, %function #endif PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_top: _PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_top: push_all Q .req w20 src0 .req x0 src1 .req x1 src2 .req x2 src3 .req x3 src4 .req x4 src5 .req x5 src6 .req x6 src7 .req x7 src8 .req x8 src9 .req x9 src10 .req x10 src11 .req x11 src12 .req x12 src13 .req x13 src14 .req x14 src15 .req x15 table .req x28 counter .req x19 ldrsh Q, [x2, #0] mov table, x1 add src0, x0, #32*0 add src1, x0, #32*1 add src2, x0, #32*2 add src3, x0, #32*3 add src4, x0, #32*4 add src5, x0, #32*5 add src6, x0, #32*6 add src7, x0, #32*7 add src8, x0, #32*8 add src9, x0, #32*9 add src10, x0, #32*10 add src11, x0, #32*11 add src12, x0, #32*12 add src13, x0, #32*13 add src14, x0, #32*14 add src15, x0, #32*15 ld1 { v0.8H, v1.8H, v2.8H, v3.8H}, [table], #64 mov v0.H[0], Q ld1 { v4.8H}, [ src0] ld1 { v5.8H}, [ src1] ld1 { v6.8H}, [ src2] ld1 { v7.8H}, [ src3] ld1 { v8.8H}, [ src4] ld1 { v9.8H}, [ src5] ld1 {v10.8H}, [ src6] ld1 {v11.8H}, [ src7] ld1 {v12.8H}, [ src8] ld1 {v13.8H}, [ src9] ld1 {v14.8H}, [src10] ld1 {v15.8H}, [src11] ld1 {v16.8H}, [src12] ld1 {v17.8H}, [src13] ld1 {v18.8H}, [src14] ld1 {v19.8H}, [src15] qo_butterfly_top v5, v7, v9, v11, v13, v15, v17, v19, v28, v29, v30, v31, v0, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3 qo_butterfly_mixed v5, v7, v9, v11, v13, v15, v17, v19, v28, v29, v30, v31, v4, v6, v8, v10, v12, v14, v16, v18, v20, v21, v22, v23, v0, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3 qo_butterfly_mixed v4, v6, v8, v10, v12, v14, v16, v18, v20, v21, v22, v23, v5, v7, v13, v15, v9, v11, v17, v19, v28, v29, v30, v31, v0, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7 qo_butterfly_mixed v5, v7, v13, v15, v9, v11, v17, v19, v28, v29, v30, v31, v4, v6, v12, v14, v8, v10, v16, v18, v20, v21, v22, v23, v0, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7 qo_butterfly_mixed v4, v6, v12, v14, v8, v10, v16, v18, v20, v21, v22, v23, v5, v9, v13, v17, v7, v11, v15, v19, v28, v29, v30, v31, v0, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7 qo_butterfly_mixed v5, v9, v13, v17, v7, v11, v15, v19, v28, v29, v30, v31, v4, v8, v12, v16, v6, v10, v14, v18, v20, v21, v22, v23, v0, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7 qo_butterfly_mixed v4, v8, v12, v16, v6, v10, v14, v18, v20, v21, v22, v23, v4, v6, v8, v10, v5, v7, v9, v11, v28, v29, v30, v31, v0, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7, v2, 0, 1, v2, 2, 3, v2, 4, 5, v2, 6, 7 qo_butterfly_mixed v4, v6, v8, v10, v5, v7, v9, v11, v28, v29, v30, v31, v12, v14, v16, v18, v13, v15, v17, v19, v20, v21, v22, v23, v0, v2, 0, 1, v2, 2, 3, v2, 4, 5, v2, 6, 7, v3, 0, 1, v3, 2, 3, v3, 4, 5, v3, 6, 7 qo_butterfly_bot v12, v14, v16, v18, v13, v15, v17, v19, v20, v21, v22, v23, v0, v3, 0, 1, v3, 2, 3, v3, 4, 5, v3, 6, 7 st1 { v4.8H}, [ src0], #16 ld1 { v4.8H}, [ src0] st1 { v5.8H}, [ src1], #16 ld1 { v5.8H}, [ src1] st1 { v6.8H}, [ src2], #16 ld1 { v6.8H}, [ src2] st1 { v7.8H}, [ src3], #16 ld1 { v7.8H}, [ src3] st1 { v8.8H}, [ src4], #16 ld1 { v8.8H}, [ src4] st1 { v9.8H}, [ src5], #16 ld1 { v9.8H}, [ src5] st1 {v10.8H}, [ src6], #16 ld1 {v10.8H}, [ src6] st1 {v11.8H}, [ src7], #16 ld1 {v11.8H}, [ src7] st1 {v12.8H}, [ src8], #16 ld1 {v12.8H}, [ src8] st1 {v13.8H}, [ src9], #16 ld1 {v13.8H}, [ src9] st1 {v14.8H}, [src10], #16 ld1 {v14.8H}, [src10] st1 {v15.8H}, [src11], #16 ld1 {v15.8H}, [src11] st1 {v16.8H}, [src12], #16 ld1 {v16.8H}, [src12] st1 {v17.8H}, [src13], #16 ld1 {v17.8H}, [src13] st1 {v18.8H}, [src14], #16 ld1 {v18.8H}, [src14] st1 {v19.8H}, [src15], #16 ld1 {v19.8H}, [src15] qo_butterfly_top v5, v7, v9, v11, v13, v15, v17, v19, v28, v29, v30, v31, v0, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3 qo_butterfly_mixed v5, v7, v9, v11, v13, v15, v17, v19, v28, v29, v30, v31, v4, v6, v8, v10, v12, v14, v16, v18, v20, v21, v22, v23, v0, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3 qo_butterfly_mixed v4, v6, v8, v10, v12, v14, v16, v18, v20, v21, v22, v23, v5, v7, v13, v15, v9, v11, v17, v19, v28, v29, v30, v31, v0, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 2, 3, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7 qo_butterfly_mixed v5, v7, v13, v15, v9, v11, v17, v19, v28, v29, v30, v31, v4, v6, v12, v14, v8, v10, v16, v18, v20, v21, v22, v23, v0, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7 qo_butterfly_mixed v4, v6, v12, v14, v8, v10, v16, v18, v20, v21, v22, v23, v5, v9, v13, v17, v7, v11, v15, v19, v28, v29, v30, v31, v0, v0, 4, 5, v0, 4, 5, v0, 6, 7, v0, 6, 7, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7 qo_butterfly_mixed v5, v9, v13, v17, v7, v11, v15, v19, v28, v29, v30, v31, v4, v8, v12, v16, v6, v10, v14, v18, v20, v21, v22, v23, v0, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7 qo_butterfly_mixed v4, v8, v12, v16, v6, v10, v14, v18, v20, v21, v22, v23, v4, v6, v8, v10, v5, v7, v9, v11, v28, v29, v30, v31, v0, v1, 0, 1, v1, 2, 3, v1, 4, 5, v1, 6, 7, v2, 0, 1, v2, 2, 3, v2, 4, 5, v2, 6, 7 qo_butterfly_mixed v4, v6, v8, v10, v5, v7, v9, v11, v28, v29, v30, v31, v12, v14, v16, v18, v13, v15, v17, v19, v20, v21, v22, v23, v0, v2, 0, 1, v2, 2, 3, v2, 4, 5, v2, 6, 7, v3, 0, 1, v3, 2, 3, v3, 4, 5, v3, 6, 7 qo_butterfly_bot v12, v14, v16, v18, v13, v15, v17, v19, v20, v21, v22, v23, v0, v3, 0, 1, v3, 2, 3, v3, 4, 5, v3, 6, 7 st1 { v4.8H}, [ src0], #16 st1 { v5.8H}, [ src1], #16 st1 { v6.8H}, [ src2], #16 st1 { v7.8H}, [ src3], #16 st1 { v8.8H}, [ src4], #16 st1 { v9.8H}, [ src5], #16 st1 {v10.8H}, [ src6], #16 st1 {v11.8H}, [ src7], #16 st1 {v12.8H}, [ src8], #16 st1 {v13.8H}, [ src9], #16 st1 {v14.8H}, [src10], #16 st1 {v15.8H}, [src11], #16 st1 {v16.8H}, [src12], #16 st1 {v17.8H}, [src13], #16 st1 {v18.8H}, [src14], #16 st1 {v19.8H}, [src15], #16 .unreq Q .unreq src0 .unreq src1 .unreq src2 .unreq src3 .unreq src4 .unreq src5 .unreq src6 .unreq src7 .unreq src8 .unreq src9 .unreq src10 .unreq src11 .unreq src12 .unreq src13 .unreq src14 .unreq src15 .unreq table .unreq counter pop_all br lr .align 2 .global PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_bot .global _PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_bot #ifndef __clang__ .type PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_bot, %function #endif PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_bot: _PQCLEAN_KYBER768_AARCH64_asm_ntt_SIMD_bot: push_all Q .req w20 BarrettM .req w21 src0 .req x0 src1 .req x1 table .req x28 counter .req x19 ldrsh Q, [x2, #0] ldrsh BarrettM, [x2, #8] add table, x1, #64 add src0, x0, #256*0 add src1, x0, #256*1 ld4 {v16.4S, v17.4S, v18.4S, v19.4S}, [src0] ld4 {v20.4S, v21.4S, v22.4S, v23.4S}, [src1] trn1 v24.4S, v16.4S, v20.4S ld2 { v0.8H, v1.8H}, [table], #32 trn2 v28.4S, v16.4S, v20.4S ld2 { v2.8H, v3.8H}, [table], #32 trn1 v25.4S, v17.4S, v21.4S ld2 { v4.8H, v5.8H}, [table], #32 trn2 v29.4S, v17.4S, v21.4S ld2 { v6.8H, v7.8H}, [table], #32 trn1 v26.4S, v18.4S, v22.4S ld2 { v8.8H, v9.8H}, [table], #32 trn2 v30.4S, v18.4S, v22.4S ld2 {v10.8H, v11.8H}, [table], #32 trn1 v27.4S, v19.4S, v23.4S ld2 {v12.8H, v13.8H}, [table], #32 trn2 v31.4S, v19.4S, v23.4S ld2 {v14.8H, v15.8H}, [table], #32 dup v0.8H, Q mov v1.H[0], BarrettM do_butterfly_vec_top v25, v27, v29, v31, v18, v19, v0, v2, v3, v2, v3 do_butterfly_vec_mixed v25, v27, v29, v31, v18, v19, v24, v26, v28, v30, v16, v17, v0, v2, v3, v2, v3, v2, v3, v2, v3 do_butterfly_vec_mixed v24, v26, v28, v30, v16, v17, v25, v29, v27, v31, v18, v19, v0, v2, v3, v2, v3, v4, v5, v6, v7 do_butterfly_vec_mixed v25, v29, v27, v31, v18, v19, v24, v28, v26, v30, v16, v17, v0, v4, v5, v6, v7, v4, v5, v6, v7 do_butterfly_vec_mixed v24, v28, v26, v30, v16, v17, v24, v26, v25, v27, v18, v19, v0, v4, v5, v6, v7, v8, v9, v10, v11 do_butterfly_vec_mixed v24, v26, v25, v27, v18, v19, v28, v30, v29, v31, v16, v17, v0, v8, v9, v10, v11, v12, v13, v14, v15 do_butterfly_vec_bot v28, v30, v29, v31, v16, v17, v0, v12, v13, v14, v15 oo_barrett v24, v25, v26, v27, v16, v17, v18, v19, v28, v29, v30, v31, v20, v21, v22, v23, v1, #11, v0 trn1 v16.4S, v24.4S, v28.4S trn2 v20.4S, v24.4S, v28.4S trn1 v17.4S, v25.4S, v29.4S trn2 v21.4S, v25.4S, v29.4S trn1 v18.4S, v26.4S, v30.4S trn2 v22.4S, v26.4S, v30.4S trn1 v19.4S, v27.4S, v31.4S trn2 v23.4S, v27.4S, v31.4S mov counter, #3 _ntt_bot_loop: st4 {v16.4S, v17.4S, v18.4S, v19.4S}, [src0], #64 ld4 {v16.4S, v17.4S, v18.4S, v19.4S}, [src0] st4 {v20.4S, v21.4S, v22.4S, v23.4S}, [src1], #64 ld4 {v20.4S, v21.4S, v22.4S, v23.4S}, [src1] trn1 v24.4S, v16.4S, v20.4S ld2 { v0.8H, v1.8H}, [table], #32 trn2 v28.4S, v16.4S, v20.4S ld2 { v2.8H, v3.8H}, [table], #32 trn1 v25.4S, v17.4S, v21.4S ld2 { v4.8H, v5.8H}, [table], #32 trn2 v29.4S, v17.4S, v21.4S ld2 { v6.8H, v7.8H}, [table], #32 trn1 v26.4S, v18.4S, v22.4S ld2 { v8.8H, v9.8H}, [table], #32 trn2 v30.4S, v18.4S, v22.4S ld2 {v10.8H, v11.8H}, [table], #32 trn1 v27.4S, v19.4S, v23.4S ld2 {v12.8H, v13.8H}, [table], #32 trn2 v31.4S, v19.4S, v23.4S ld2 {v14.8H, v15.8H}, [table], #32 dup v0.8H, Q mov v1.H[0], BarrettM do_butterfly_vec_top v25, v27, v29, v31, v18, v19, v0, v2, v3, v2, v3 do_butterfly_vec_mixed v25, v27, v29, v31, v18, v19, v24, v26, v28, v30, v16, v17, v0, v2, v3, v2, v3, v2, v3, v2, v3 do_butterfly_vec_mixed v24, v26, v28, v30, v16, v17, v25, v29, v27, v31, v18, v19, v0, v2, v3, v2, v3, v4, v5, v6, v7 do_butterfly_vec_mixed v25, v29, v27, v31, v18, v19, v24, v28, v26, v30, v16, v17, v0, v4, v5, v6, v7, v4, v5, v6, v7 do_butterfly_vec_mixed v24, v28, v26, v30, v16, v17, v24, v26, v25, v27, v18, v19, v0, v4, v5, v6, v7, v8, v9, v10, v11 do_butterfly_vec_mixed v24, v26, v25, v27, v18, v19, v28, v30, v29, v31, v16, v17, v0, v8, v9, v10, v11, v12, v13, v14, v15 do_butterfly_vec_bot v28, v30, v29, v31, v16, v17, v0, v12, v13, v14, v15 oo_barrett v24, v25, v26, v27, v16, v17, v18, v19, v28, v29, v30, v31, v20, v21, v22, v23, v1, #11, v0 trn1 v16.4S, v24.4S, v28.4S trn2 v20.4S, v24.4S, v28.4S trn1 v17.4S, v25.4S, v29.4S trn2 v21.4S, v25.4S, v29.4S trn1 v18.4S, v26.4S, v30.4S trn2 v22.4S, v26.4S, v30.4S trn1 v19.4S, v27.4S, v31.4S trn2 v23.4S, v27.4S, v31.4S sub counter, counter, #1 cbnz counter, _ntt_bot_loop st4 {v16.4S, v17.4S, v18.4S, v19.4S}, [src0], #64 st4 {v20.4S, v21.4S, v22.4S, v23.4S}, [src1], #64 .unreq Q .unreq BarrettM .unreq src0 .unreq src1 .unreq table .unreq counter pop_all br lr