diff options
Diffstat (limited to 'target/arm/tcg/vec_helper.c')
| -rw-r--r-- | target/arm/tcg/vec_helper.c | 384 |
1 files changed, 331 insertions, 53 deletions
diff --git a/target/arm/tcg/vec_helper.c b/target/arm/tcg/vec_helper.c index 986eaf8..0603db0 100644 --- a/target/arm/tcg/vec_helper.c +++ b/target/arm/tcg/vec_helper.c @@ -825,11 +825,11 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \ clear_tail(d, opr_sz, simd_maxsz(desc)); \ } -DO_DOT(gvec_sdot_b, int32_t, int8_t, int8_t) -DO_DOT(gvec_udot_b, uint32_t, uint8_t, uint8_t) -DO_DOT(gvec_usdot_b, uint32_t, uint8_t, int8_t) -DO_DOT(gvec_sdot_h, int64_t, int16_t, int16_t) -DO_DOT(gvec_udot_h, uint64_t, uint16_t, uint16_t) +DO_DOT(gvec_sdot_4b, int32_t, int8_t, int8_t) +DO_DOT(gvec_udot_4b, uint32_t, uint8_t, uint8_t) +DO_DOT(gvec_usdot_4b, uint32_t, uint8_t, int8_t) +DO_DOT(gvec_sdot_4h, int64_t, int16_t, int16_t) +DO_DOT(gvec_udot_4h, uint64_t, uint16_t, uint16_t) #define DO_DOT_IDX(NAME, TYPED, TYPEN, TYPEM, HD) \ void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \ @@ -865,12 +865,63 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \ clear_tail(d, opr_sz, simd_maxsz(desc)); \ } -DO_DOT_IDX(gvec_sdot_idx_b, int32_t, int8_t, int8_t, H4) -DO_DOT_IDX(gvec_udot_idx_b, uint32_t, uint8_t, uint8_t, H4) -DO_DOT_IDX(gvec_sudot_idx_b, int32_t, int8_t, uint8_t, H4) -DO_DOT_IDX(gvec_usdot_idx_b, int32_t, uint8_t, int8_t, H4) -DO_DOT_IDX(gvec_sdot_idx_h, int64_t, int16_t, int16_t, H8) -DO_DOT_IDX(gvec_udot_idx_h, uint64_t, uint16_t, uint16_t, H8) +DO_DOT_IDX(gvec_sdot_idx_4b, int32_t, int8_t, int8_t, H4) +DO_DOT_IDX(gvec_udot_idx_4b, uint32_t, uint8_t, uint8_t, H4) +DO_DOT_IDX(gvec_sudot_idx_4b, int32_t, int8_t, uint8_t, H4) +DO_DOT_IDX(gvec_usdot_idx_4b, int32_t, uint8_t, int8_t, H4) +DO_DOT_IDX(gvec_sdot_idx_4h, int64_t, int16_t, int16_t, H8) +DO_DOT_IDX(gvec_udot_idx_4h, uint64_t, uint16_t, uint16_t, H8) + +#undef DO_DOT +#undef DO_DOT_IDX + +/* Similar for 2-way dot product */ +#define DO_DOT(NAME, TYPED, TYPEN, TYPEM) \ +void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \ +{ \ + intptr_t i, opr_sz = simd_oprsz(desc); \ + TYPED *d = vd, *a = va; \ + TYPEN *n = vn; \ + TYPEM *m = vm; \ + for (i = 0; i < opr_sz / sizeof(TYPED); ++i) { \ + d[i] = (a[i] + \ + (TYPED)n[i * 2 + 0] * m[i * 2 + 0] + \ + (TYPED)n[i * 2 + 1] * m[i * 2 + 1]); \ + } \ + clear_tail(d, opr_sz, simd_maxsz(desc)); \ +} + +#define DO_DOT_IDX(NAME, TYPED, TYPEN, TYPEM, HD) \ +void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, uint32_t desc) \ +{ \ + intptr_t i = 0, opr_sz = simd_oprsz(desc); \ + intptr_t opr_sz_n = opr_sz / sizeof(TYPED); \ + intptr_t segend = MIN(16 / sizeof(TYPED), opr_sz_n); \ + intptr_t index = simd_data(desc); \ + TYPED *d = vd, *a = va; \ + TYPEN *n = vn; \ + TYPEM *m_indexed = (TYPEM *)vm + HD(index) * 2; \ + do { \ + TYPED m0 = m_indexed[i * 2 + 0]; \ + TYPED m1 = m_indexed[i * 2 + 1]; \ + do { \ + d[i] = (a[i] + \ + n[i * 2 + 0] * m0 + \ + n[i * 2 + 1] * m1); \ + } while (++i < segend); \ + segend = i + (16 / sizeof(TYPED)); \ + } while (i < opr_sz_n); \ + clear_tail(d, opr_sz, simd_maxsz(desc)); \ +} + +DO_DOT(gvec_sdot_2h, int32_t, int16_t, int16_t) +DO_DOT(gvec_udot_2h, uint32_t, uint16_t, uint16_t) + +DO_DOT_IDX(gvec_sdot_idx_2h, int32_t, int16_t, int16_t, H4) +DO_DOT_IDX(gvec_udot_idx_2h, uint32_t, uint16_t, uint16_t, H4) + +#undef DO_DOT +#undef DO_DOT_IDX void HELPER(gvec_fcaddh)(void *vd, void *vn, void *vm, float_status *fpst, uint32_t desc) @@ -1419,10 +1470,12 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, \ DO_3OP(gvec_fadd_h, float16_add, float16) DO_3OP(gvec_fadd_s, float32_add, float32) DO_3OP(gvec_fadd_d, float64_add, float64) +DO_3OP(gvec_bfadd, bfloat16_add, bfloat16) DO_3OP(gvec_fsub_h, float16_sub, float16) DO_3OP(gvec_fsub_s, float32_sub, float32) DO_3OP(gvec_fsub_d, float64_sub, float64) +DO_3OP(gvec_bfsub, bfloat16_sub, bfloat16) DO_3OP(gvec_fmul_h, float16_mul, float16) DO_3OP(gvec_fmul_s, float32_mul, float32) @@ -1515,6 +1568,13 @@ DO_3OP(gvec_ah_fmin_h, helper_vfp_ah_minh, float16) DO_3OP(gvec_ah_fmin_s, helper_vfp_ah_mins, float32) DO_3OP(gvec_ah_fmin_d, helper_vfp_ah_mind, float64) +DO_3OP(gvec_fmax_b16, bfloat16_max, bfloat16) +DO_3OP(gvec_fmin_b16, bfloat16_min, bfloat16) +DO_3OP(gvec_fmaxnum_b16, bfloat16_maxnum, bfloat16) +DO_3OP(gvec_fminnum_b16, bfloat16_minnum, bfloat16) +DO_3OP(gvec_ah_fmax_b16, helper_sme2_ah_fmax_b16, bfloat16) +DO_3OP(gvec_ah_fmin_b16, helper_sme2_ah_fmin_b16, bfloat16) + #endif #undef DO_3OP @@ -1550,6 +1610,12 @@ static float16 float16_muladd_f(float16 dest, float16 op1, float16 op2, return float16_muladd(op1, op2, dest, 0, stat); } +static bfloat16 bfloat16_muladd_f(bfloat16 dest, bfloat16 op1, bfloat16 op2, + float_status *stat) +{ + return bfloat16_muladd(op1, op2, dest, 0, stat); +} + static float32 float32_muladd_f(float32 dest, float32 op1, float32 op2, float_status *stat) { @@ -1568,6 +1634,12 @@ static float16 float16_mulsub_f(float16 dest, float16 op1, float16 op2, return float16_muladd(float16_chs(op1), op2, dest, 0, stat); } +static bfloat16 bfloat16_mulsub_f(bfloat16 dest, bfloat16 op1, bfloat16 op2, + float_status *stat) +{ + return bfloat16_muladd(bfloat16_chs(op1), op2, dest, 0, stat); +} + static float32 float32_mulsub_f(float32 dest, float32 op1, float32 op2, float_status *stat) { @@ -1586,6 +1658,12 @@ static float16 float16_ah_mulsub_f(float16 dest, float16 op1, float16 op2, return float16_muladd(op1, op2, dest, float_muladd_negate_product, stat); } +static bfloat16 bfloat16_ah_mulsub_f(bfloat16 dest, bfloat16 op1, bfloat16 op2, + float_status *stat) +{ + return bfloat16_muladd(op1, op2, dest, float_muladd_negate_product, stat); +} + static float32 float32_ah_mulsub_f(float32 dest, float32 op1, float32 op2, float_status *stat) { @@ -1610,23 +1688,28 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, \ clear_tail(d, oprsz, simd_maxsz(desc)); \ } -DO_MULADD(gvec_fmla_h, float16_muladd_nf, float16) -DO_MULADD(gvec_fmla_s, float32_muladd_nf, float32) +DO_MULADD(gvec_fmla_nf_h, float16_muladd_nf, float16) +DO_MULADD(gvec_fmla_nf_s, float32_muladd_nf, float32) -DO_MULADD(gvec_fmls_h, float16_mulsub_nf, float16) -DO_MULADD(gvec_fmls_s, float32_mulsub_nf, float32) +DO_MULADD(gvec_fmls_nf_h, float16_mulsub_nf, float16) +DO_MULADD(gvec_fmls_nf_s, float32_mulsub_nf, float32) DO_MULADD(gvec_vfma_h, float16_muladd_f, float16) DO_MULADD(gvec_vfma_s, float32_muladd_f, float32) DO_MULADD(gvec_vfma_d, float64_muladd_f, float64) +DO_MULADD(gvec_bfmla, bfloat16_muladd_f, bfloat16) DO_MULADD(gvec_vfms_h, float16_mulsub_f, float16) DO_MULADD(gvec_vfms_s, float32_mulsub_f, float32) DO_MULADD(gvec_vfms_d, float64_mulsub_f, float64) +DO_MULADD(gvec_bfmls, bfloat16_mulsub_f, bfloat16) DO_MULADD(gvec_ah_vfms_h, float16_ah_mulsub_f, float16) DO_MULADD(gvec_ah_vfms_s, float32_ah_mulsub_f, float32) DO_MULADD(gvec_ah_vfms_d, float64_ah_mulsub_f, float64) +DO_MULADD(gvec_ah_bfmls, bfloat16_ah_mulsub_f, bfloat16) + +#undef DO_MULADD /* For the indexed ops, SVE applies the index per 128-bit vector segment. * For AdvSIMD, there is of course only one such vector segment. @@ -1745,14 +1828,17 @@ void HELPER(NAME)(void *vd, void *vn, void *vm, void *va, \ DO_FMLA_IDX(gvec_fmla_idx_h, float16, H2, 0, 0) DO_FMLA_IDX(gvec_fmla_idx_s, float32, H4, 0, 0) DO_FMLA_IDX(gvec_fmla_idx_d, float64, H8, 0, 0) +DO_FMLA_IDX(gvec_bfmla_idx, bfloat16, H2, 0, 0) DO_FMLA_IDX(gvec_fmls_idx_h, float16, H2, INT16_MIN, 0) DO_FMLA_IDX(gvec_fmls_idx_s, float32, H4, INT32_MIN, 0) DO_FMLA_IDX(gvec_fmls_idx_d, float64, H8, INT64_MIN, 0) +DO_FMLA_IDX(gvec_bfmls_idx, bfloat16, H2, INT16_MIN, 0) DO_FMLA_IDX(gvec_ah_fmls_idx_h, float16, H2, 0, float_muladd_negate_product) DO_FMLA_IDX(gvec_ah_fmls_idx_s, float32, H4, 0, float_muladd_negate_product) DO_FMLA_IDX(gvec_ah_fmls_idx_d, float64, H8, 0, float_muladd_negate_product) +DO_FMLA_IDX(gvec_ah_bfmls_idx, bfloat16, H2, 0, float_muladd_negate_product) #undef DO_FMLA_IDX @@ -2184,7 +2270,8 @@ void HELPER(sve2_fmlal_zzzw_s)(void *vd, void *vn, void *vm, void *va, intptr_t i, oprsz = simd_oprsz(desc); bool is_s = extract32(desc, SIMD_DATA_SHIFT, 1); intptr_t sel = extract32(desc, SIMD_DATA_SHIFT + 1, 1) * sizeof(float16); - float_status *status = &env->vfp.fp_status[FPST_A64]; + bool za = extract32(desc, SIMD_DATA_SHIFT + 2, 1); + float_status *status = &env->vfp.fp_status[za ? FPST_ZA : FPST_A64]; bool fz16 = env->vfp.fpcr & FPCR_FZ16; int negx = 0, negf = 0; @@ -2267,8 +2354,9 @@ void HELPER(sve2_fmlal_zzxw_s)(void *vd, void *vn, void *vm, void *va, intptr_t i, j, oprsz = simd_oprsz(desc); bool is_s = extract32(desc, SIMD_DATA_SHIFT, 1); intptr_t sel = extract32(desc, SIMD_DATA_SHIFT + 1, 1) * sizeof(float16); - intptr_t idx = extract32(desc, SIMD_DATA_SHIFT + 2, 3) * sizeof(float16); - float_status *status = &env->vfp.fp_status[FPST_A64]; + bool za = extract32(desc, SIMD_DATA_SHIFT + 2, 1); + intptr_t idx = extract32(desc, SIMD_DATA_SHIFT + 3, 3) * sizeof(float16); + float_status *status = &env->vfp.fp_status[za ? FPST_ZA : FPST_A64]; bool fz16 = env->vfp.fpcr & FPCR_FZ16; int negx = 0, negf = 0; @@ -2989,31 +3077,62 @@ float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2, float_status *fpst) float32 bfdotadd_ebf(float32 sum, uint32_t e1, uint32_t e2, float_status *fpst, float_status *fpst_odd) { - /* - * Compare f16_dotadd() in sme_helper.c, but here we have - * bfloat16 inputs. In particular that means that we do not - * want the FPCR.FZ16 flush semantics, so we use the normal - * float_status for the input handling here. - */ - float64 e1r = float32_to_float64(e1 << 16, fpst); - float64 e1c = float32_to_float64(e1 & 0xffff0000u, fpst); - float64 e2r = float32_to_float64(e2 << 16, fpst); - float64 e2c = float32_to_float64(e2 & 0xffff0000u, fpst); - float64 t64; + float32 s1r = e1 << 16; + float32 s1c = e1 & 0xffff0000u; + float32 s2r = e2 << 16; + float32 s2c = e2 & 0xffff0000u; float32 t32; - /* - * The ARM pseudocode function FPDot performs both multiplies - * and the add with a single rounding operation. Emulate this - * by performing the first multiply in round-to-odd, then doing - * the second multiply as fused multiply-add, and rounding to - * float32 all in one step. - */ - t64 = float64_mul(e1r, e2r, fpst_odd); - t64 = float64r32_muladd(e1c, e2c, t64, 0, fpst); + /* C.f. FPProcessNaNs4 */ + if (float32_is_any_nan(s1r) || float32_is_any_nan(s1c) || + float32_is_any_nan(s2r) || float32_is_any_nan(s2c)) { + if (float32_is_signaling_nan(s1r, fpst)) { + t32 = s1r; + } else if (float32_is_signaling_nan(s1c, fpst)) { + t32 = s1c; + } else if (float32_is_signaling_nan(s2r, fpst)) { + t32 = s2r; + } else if (float32_is_signaling_nan(s2c, fpst)) { + t32 = s2c; + } else if (float32_is_any_nan(s1r)) { + t32 = s1r; + } else if (float32_is_any_nan(s1c)) { + t32 = s1c; + } else if (float32_is_any_nan(s2r)) { + t32 = s2r; + } else { + t32 = s2c; + } + /* + * FPConvertNaN(FPProcessNaN(t32)) will be done as part + * of the final addition below. + */ + } else { + /* + * Compare f16_dotadd() in sme_helper.c, but here we have + * bfloat16 inputs. In particular that means that we do not + * want the FPCR.FZ16 flush semantics, so we use the normal + * float_status for the input handling here. + */ + float64 e1r = float32_to_float64(s1r, fpst); + float64 e1c = float32_to_float64(s1c, fpst); + float64 e2r = float32_to_float64(s2r, fpst); + float64 e2c = float32_to_float64(s2c, fpst); + float64 t64; + + /* + * The ARM pseudocode function FPDot performs both multiplies + * and the add with a single rounding operation. Emulate this + * by performing the first multiply in round-to-odd, then doing + * the second multiply as fused multiply-add, and rounding to + * float32 all in one step. + */ + t64 = float64_mul(e1r, e2r, fpst_odd); + t64 = float64r32_muladd(e1c, e2c, t64, 0, fpst); - /* This conversion is exact, because we've already rounded. */ - t32 = float64_to_float32(t64, fpst); + /* This conversion is exact, because we've already rounded. */ + t32 = float64_to_float32(t64, fpst); + } /* The final accumulation step is not fused. */ return float32_add(sum, t32, fpst); @@ -3070,6 +3189,45 @@ void HELPER(gvec_bfdot_idx)(void *vd, void *vn, void *vm, clear_tail(d, opr_sz, simd_maxsz(desc)); } +void HELPER(sme2_bfvdot_idx)(void *vd, void *vn, void *vm, + void *va, CPUARMState *env, uint32_t desc) +{ + intptr_t i, j, opr_sz = simd_oprsz(desc); + intptr_t idx = extract32(desc, SIMD_DATA_SHIFT, 2); + intptr_t sel = extract32(desc, SIMD_DATA_SHIFT + 2, 1); + intptr_t elements = opr_sz / 4; + intptr_t eltspersegment = MIN(16 / 4, elements); + float32 *d = vd, *a = va; + uint16_t *n0 = vn; + uint16_t *n1 = vn + sizeof(ARMVectorReg); + uint32_t *m = vm; + float_status fpst, fpst_odd; + + if (is_ebf(env, &fpst, &fpst_odd)) { + for (i = 0; i < elements; i += eltspersegment) { + uint32_t m_idx = m[i + H4(idx)]; + + for (j = 0; j < eltspersegment; j++) { + uint32_t nn = (n0[H2(2 * (i + j) + sel)]) + | (n1[H2(2 * (i + j) + sel)] << 16); + d[i + H4(j)] = bfdotadd_ebf(a[i + H4(j)], nn, m_idx, + &fpst, &fpst_odd); + } + } + } else { + for (i = 0; i < elements; i += eltspersegment) { + uint32_t m_idx = m[i + H4(idx)]; + + for (j = 0; j < eltspersegment; j++) { + uint32_t nn = (n0[H2(2 * (i + j) + sel)]) + | (n1[H2(2 * (i + j) + sel)] << 16); + d[i + H4(j)] = bfdotadd(a[i + H4(j)], nn, m_idx, &fpst); + } + } + } + clear_tail(d, opr_sz, simd_maxsz(desc)); +} + void HELPER(gvec_bfmmla)(void *vd, void *vn, void *vm, void *va, CPUARMState *env, uint32_t desc) { @@ -3146,44 +3304,76 @@ void HELPER(gvec_bfmmla)(void *vd, void *vn, void *vm, void *va, clear_tail(d, opr_sz, simd_maxsz(desc)); } -void HELPER(gvec_bfmlal)(void *vd, void *vn, void *vm, void *va, - float_status *stat, uint32_t desc) +static void do_bfmlal(float32 *d, bfloat16 *n, bfloat16 *m, float32 *a, + float_status *stat, uint32_t desc, int negx, int negf) { intptr_t i, opr_sz = simd_oprsz(desc); - intptr_t sel = simd_data(desc); - float32 *d = vd, *a = va; - bfloat16 *n = vn, *m = vm; + intptr_t sel = extract32(desc, SIMD_DATA_SHIFT, 1); for (i = 0; i < opr_sz / 4; ++i) { - float32 nn = n[H2(i * 2 + sel)] << 16; + float32 nn = (negx ^ n[H2(i * 2 + sel)]) << 16; float32 mm = m[H2(i * 2 + sel)] << 16; - d[H4(i)] = float32_muladd(nn, mm, a[H4(i)], 0, stat); + d[H4(i)] = float32_muladd(nn, mm, a[H4(i)], negf, stat); } clear_tail(d, opr_sz, simd_maxsz(desc)); } -void HELPER(gvec_bfmlal_idx)(void *vd, void *vn, void *vm, - void *va, float_status *stat, uint32_t desc) +void HELPER(gvec_bfmlal)(void *vd, void *vn, void *vm, void *va, + float_status *stat, uint32_t desc) +{ + do_bfmlal(vd, vn, vm, va, stat, desc, 0, 0); +} + +void HELPER(gvec_bfmlsl)(void *vd, void *vn, void *vm, void *va, + float_status *stat, uint32_t desc) +{ + do_bfmlal(vd, vn, vm, va, stat, desc, 0x8000, 0); +} + +void HELPER(gvec_ah_bfmlsl)(void *vd, void *vn, void *vm, void *va, + float_status *stat, uint32_t desc) +{ + do_bfmlal(vd, vn, vm, va, stat, desc, 0, float_muladd_negate_product); +} + +static void do_bfmlal_idx(float32 *d, bfloat16 *n, bfloat16 *m, float32 *a, + float_status *stat, uint32_t desc, int negx, int negf) { intptr_t i, j, opr_sz = simd_oprsz(desc); intptr_t sel = extract32(desc, SIMD_DATA_SHIFT, 1); intptr_t index = extract32(desc, SIMD_DATA_SHIFT + 1, 3); intptr_t elements = opr_sz / 4; intptr_t eltspersegment = MIN(16 / 4, elements); - float32 *d = vd, *a = va; - bfloat16 *n = vn, *m = vm; for (i = 0; i < elements; i += eltspersegment) { float32 m_idx = m[H2(2 * i + index)] << 16; for (j = i; j < i + eltspersegment; j++) { - float32 n_j = n[H2(2 * j + sel)] << 16; - d[H4(j)] = float32_muladd(n_j, m_idx, a[H4(j)], 0, stat); + float32 n_j = (negx ^ n[H2(2 * j + sel)]) << 16; + d[H4(j)] = float32_muladd(n_j, m_idx, a[H4(j)], negf, stat); } } clear_tail(d, opr_sz, simd_maxsz(desc)); } +void HELPER(gvec_bfmlal_idx)(void *vd, void *vn, void *vm, void *va, + float_status *stat, uint32_t desc) +{ + do_bfmlal_idx(vd, vn, vm, va, stat, desc, 0, 0); +} + +void HELPER(gvec_bfmlsl_idx)(void *vd, void *vn, void *vm, void *va, + float_status *stat, uint32_t desc) +{ + do_bfmlal_idx(vd, vn, vm, va, stat, desc, 0x8000, 0); +} + +void HELPER(gvec_ah_bfmlsl_idx)(void *vd, void *vn, void *vm, void *va, + float_status *stat, uint32_t desc) +{ + do_bfmlal_idx(vd, vn, vm, va, stat, desc, 0, float_muladd_negate_product); +} + #define DO_CLAMP(NAME, TYPE) \ void HELPER(NAME)(void *d, void *n, void *m, void *a, uint32_t desc) \ { \ @@ -3253,3 +3443,91 @@ void HELPER(gvec_ursqrte_s)(void *vd, void *vn, uint32_t desc) } clear_tail(d, opr_sz, simd_maxsz(desc)); } + +static inline void do_lut_b(void *zd, uint64_t *indexes, uint64_t *table, + unsigned elements, unsigned segbase, + unsigned dstride, unsigned isize, + unsigned tsize, unsigned nreg) +{ + for (unsigned r = 0; r < nreg; ++r) { + uint8_t *dst = zd + dstride * r; + unsigned base = segbase + r * elements; + + for (unsigned e = 0; e < elements; ++e) { + unsigned index = extractn(indexes, (base + e) * isize, isize); + dst[H1(e)] = extractn(table, index * tsize, 8); + } + } +} + +static inline void do_lut_h(void *zd, uint64_t *indexes, uint64_t *table, + unsigned elements, unsigned segbase, + unsigned dstride, unsigned isize, + unsigned tsize, unsigned nreg) +{ + for (unsigned r = 0; r < nreg; ++r) { + uint16_t *dst = zd + dstride * r; + unsigned base = segbase + r * elements; + + for (unsigned e = 0; e < elements; ++e) { + unsigned index = extractn(indexes, (base + e) * isize, isize); + dst[H2(e)] = extractn(table, index * tsize, 16); + } + } +} + +static inline void do_lut_s(void *zd, uint64_t *indexes, uint32_t *table, + unsigned elements, unsigned segbase, + unsigned dstride, unsigned isize, + unsigned tsize, unsigned nreg) +{ + for (unsigned r = 0; r < nreg; ++r) { + uint32_t *dst = zd + dstride * r; + unsigned base = segbase + r * elements; + + for (unsigned e = 0; e < elements; ++e) { + unsigned index = extractn(indexes, (base + e) * isize, isize); + dst[H4(e)] = table[H4(index)]; + } + } +} + +#define DO_SME2_LUT(ISIZE, NREG, SUFF, ESIZE) \ +void helper_sme2_luti##ISIZE##_##NREG##SUFF \ + (void *zd, void *zn, CPUARMState *env, uint32_t desc) \ +{ \ + unsigned vl = simd_oprsz(desc); \ + unsigned strided = extract32(desc, SIMD_DATA_SHIFT, 1); \ + unsigned idx = extract32(desc, SIMD_DATA_SHIFT + 1, 4); \ + unsigned elements = vl / ESIZE; \ + unsigned dstride = (!strided ? 1 : NREG == 4 ? 4 : 8); \ + unsigned segments = (ESIZE * 8) / (ISIZE * NREG); \ + unsigned segment = idx & (segments - 1); \ + ARMVectorReg indexes; \ + memcpy(&indexes, zn, vl); \ + do_lut_##SUFF(zd, indexes.d, (void *)env->za_state.zt0, elements, \ + segment * NREG * elements, \ + dstride * sizeof(ARMVectorReg), ISIZE, 32, NREG); \ +} + +DO_SME2_LUT(2,1,b, 1) +DO_SME2_LUT(2,1,h, 2) +DO_SME2_LUT(2,1,s, 4) +DO_SME2_LUT(2,2,b, 1) +DO_SME2_LUT(2,2,h, 2) +DO_SME2_LUT(2,2,s, 4) +DO_SME2_LUT(2,4,b, 1) +DO_SME2_LUT(2,4,h, 2) +DO_SME2_LUT(2,4,s, 4) + +DO_SME2_LUT(4,1,b, 1) +DO_SME2_LUT(4,1,h, 2) +DO_SME2_LUT(4,1,s, 4) +DO_SME2_LUT(4,2,b, 1) +DO_SME2_LUT(4,2,h, 2) +DO_SME2_LUT(4,2,s, 4) +DO_SME2_LUT(4,4,b, 1) +DO_SME2_LUT(4,4,h, 2) +DO_SME2_LUT(4,4,s, 4) + +#undef DO_SME2_LUT |