[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH v10 70/77] target/riscv: rvv-1.0: floating-point reciprocal e
From: |
Alistair Francis |
Subject: |
Re: [PATCH v10 70/77] target/riscv: rvv-1.0: floating-point reciprocal estimate instruction |
Date: |
Wed, 8 Dec 2021 13:38:37 +1000 |
On Mon, Nov 29, 2021 at 1:58 PM <frank.chang@sifive.com> wrote:
>
> From: Frank Chang <frank.chang@sifive.com>
>
> Implement the floating-point reciprocal estimate to 7 bits instruction.
>
> Signed-off-by: Frank Chang <frank.chang@sifive.com>
Acked-by: Alistair Francis <alistair.francis@wdc.com>
Alistair
> ---
> target/riscv/helper.h | 4 +
> target/riscv/insn32.decode | 1 +
> target/riscv/insn_trans/trans_rvv.c.inc | 1 +
> target/riscv/vector_helper.c | 191 ++++++++++++++++++++++++
> 4 files changed, 197 insertions(+)
>
> diff --git a/target/riscv/helper.h b/target/riscv/helper.h
> index bdf06dfb24d..ab283d12b79 100644
> --- a/target/riscv/helper.h
> +++ b/target/riscv/helper.h
> @@ -845,6 +845,10 @@ DEF_HELPER_5(vfrsqrt7_v_h, void, ptr, ptr, ptr, env, i32)
> DEF_HELPER_5(vfrsqrt7_v_w, void, ptr, ptr, ptr, env, i32)
> DEF_HELPER_5(vfrsqrt7_v_d, void, ptr, ptr, ptr, env, i32)
>
> +DEF_HELPER_5(vfrec7_v_h, void, ptr, ptr, ptr, env, i32)
> +DEF_HELPER_5(vfrec7_v_w, void, ptr, ptr, ptr, env, i32)
> +DEF_HELPER_5(vfrec7_v_d, void, ptr, ptr, ptr, env, i32)
> +
> DEF_HELPER_6(vfmin_vv_h, void, ptr, ptr, ptr, ptr, env, i32)
> DEF_HELPER_6(vfmin_vv_w, void, ptr, ptr, ptr, ptr, env, i32)
> DEF_HELPER_6(vfmin_vv_d, void, ptr, ptr, ptr, ptr, env, i32)
> diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode
> index 6e5f288943a..952768f8ded 100644
> --- a/target/riscv/insn32.decode
> +++ b/target/riscv/insn32.decode
> @@ -561,6 +561,7 @@ vfwnmsac_vv 111111 . ..... ..... 001 ..... 1010111
> @r_vm
> vfwnmsac_vf 111111 . ..... ..... 101 ..... 1010111 @r_vm
> vfsqrt_v 010011 . ..... 00000 001 ..... 1010111 @r2_vm
> vfrsqrt7_v 010011 . ..... 00100 001 ..... 1010111 @r2_vm
> +vfrec7_v 010011 . ..... 00101 001 ..... 1010111 @r2_vm
> vfmin_vv 000100 . ..... ..... 001 ..... 1010111 @r_vm
> vfmin_vf 000100 . ..... ..... 101 ..... 1010111 @r_vm
> vfmax_vv 000110 . ..... ..... 001 ..... 1010111 @r_vm
> diff --git a/target/riscv/insn_trans/trans_rvv.c.inc
> b/target/riscv/insn_trans/trans_rvv.c.inc
> index 8fe718610a9..ff8f6df8f7b 100644
> --- a/target/riscv/insn_trans/trans_rvv.c.inc
> +++ b/target/riscv/insn_trans/trans_rvv.c.inc
> @@ -2408,6 +2408,7 @@ static bool trans_##NAME(DisasContext *s, arg_rmr *a) \
>
> GEN_OPFV_TRANS(vfsqrt_v, opfv_check, RISCV_FRM_DYN)
> GEN_OPFV_TRANS(vfrsqrt7_v, opfv_check, RISCV_FRM_DYN)
> +GEN_OPFV_TRANS(vfrec7_v, opfv_check, RISCV_FRM_DYN)
>
> /* Vector Floating-Point MIN/MAX Instructions */
> GEN_OPFVV_TRANS(vfmin_vv, opfvv_check)
> diff --git a/target/riscv/vector_helper.c b/target/riscv/vector_helper.c
> index d5f3229bcb4..946dca53ffd 100644
> --- a/target/riscv/vector_helper.c
> +++ b/target/riscv/vector_helper.c
> @@ -3587,6 +3587,197 @@ GEN_VEXT_V_ENV(vfrsqrt7_v_h, 2, 2)
> GEN_VEXT_V_ENV(vfrsqrt7_v_w, 4, 4)
> GEN_VEXT_V_ENV(vfrsqrt7_v_d, 8, 8)
>
> +/*
> + * Vector Floating-Point Reciprocal Estimate Instruction
> + *
> + * Adapted from riscv-v-spec recip.c:
> + * https://github.com/riscv/riscv-v-spec/blob/master/recip.c
> + */
> +static uint64_t frec7(uint64_t f, int exp_size, int frac_size,
> + float_status *s)
> +{
> + uint64_t sign = extract64(f, frac_size + exp_size, 1);
> + uint64_t exp = extract64(f, frac_size, exp_size);
> + uint64_t frac = extract64(f, 0, frac_size);
> +
> + const uint8_t lookup_table[] = {
> + 127, 125, 123, 121, 119, 117, 116, 114,
> + 112, 110, 109, 107, 105, 104, 102, 100,
> + 99, 97, 96, 94, 93, 91, 90, 88,
> + 87, 85, 84, 83, 81, 80, 79, 77,
> + 76, 75, 74, 72, 71, 70, 69, 68,
> + 66, 65, 64, 63, 62, 61, 60, 59,
> + 58, 57, 56, 55, 54, 53, 52, 51,
> + 50, 49, 48, 47, 46, 45, 44, 43,
> + 42, 41, 40, 40, 39, 38, 37, 36,
> + 35, 35, 34, 33, 32, 31, 31, 30,
> + 29, 28, 28, 27, 26, 25, 25, 24,
> + 23, 23, 22, 21, 21, 20, 19, 19,
> + 18, 17, 17, 16, 15, 15, 14, 14,
> + 13, 12, 12, 11, 11, 10, 9, 9,
> + 8, 8, 7, 7, 6, 5, 5, 4,
> + 4, 3, 3, 2, 2, 1, 1, 0
> + };
> + const int precision = 7;
> +
> + if (exp == 0 && frac != 0) { /* subnormal */
> + /* Normalize the subnormal. */
> + while (extract64(frac, frac_size - 1, 1) == 0) {
> + exp--;
> + frac <<= 1;
> + }
> +
> + frac = (frac << 1) & MAKE_64BIT_MASK(0, frac_size);
> +
> + if (exp != 0 && exp != UINT64_MAX) {
> + /*
> + * Overflow to inf or max value of same sign,
> + * depending on sign and rounding mode.
> + */
> + s->float_exception_flags |= (float_flag_inexact |
> + float_flag_overflow);
> +
> + if ((s->float_rounding_mode == float_round_to_zero) ||
> + ((s->float_rounding_mode == float_round_down) && !sign) ||
> + ((s->float_rounding_mode == float_round_up) && sign)) {
> + /* Return greatest/negative finite value. */
> + return (sign << (exp_size + frac_size)) |
> + (MAKE_64BIT_MASK(frac_size, exp_size) - 1);
> + } else {
> + /* Return +-inf. */
> + return (sign << (exp_size + frac_size)) |
> + MAKE_64BIT_MASK(frac_size, exp_size);
> + }
> + }
> + }
> +
> + int idx = frac >> (frac_size - precision);
> + uint64_t out_frac = (uint64_t)(lookup_table[idx]) <<
> + (frac_size - precision);
> + uint64_t out_exp = 2 * MAKE_64BIT_MASK(0, exp_size - 1) + ~exp;
> +
> + if (out_exp == 0 || out_exp == UINT64_MAX) {
> + /*
> + * The result is subnormal, but don't raise the underflow exception,
> + * because there's no additional loss of precision.
> + */
> + out_frac = (out_frac >> 1) | MAKE_64BIT_MASK(frac_size - 1, 1);
> + if (out_exp == UINT64_MAX) {
> + out_frac >>= 1;
> + out_exp = 0;
> + }
> + }
> +
> + uint64_t val = 0;
> + val = deposit64(val, 0, frac_size, out_frac);
> + val = deposit64(val, frac_size, exp_size, out_exp);
> + val = deposit64(val, frac_size + exp_size, 1, sign);
> + return val;
> +}
> +
> +static float16 frec7_h(float16 f, float_status *s)
> +{
> + int exp_size = 5, frac_size = 10;
> + bool sign = float16_is_neg(f);
> +
> + /* frec7(+-inf) = +-0 */
> + if (float16_is_infinity(f)) {
> + return float16_set_sign(float16_zero, sign);
> + }
> +
> + /* frec7(+-0) = +-inf */
> + if (float16_is_zero(f)) {
> + s->float_exception_flags |= float_flag_divbyzero;
> + return float16_set_sign(float16_infinity, sign);
> + }
> +
> + /* frec7(sNaN) = canonical NaN */
> + if (float16_is_signaling_nan(f, s)) {
> + s->float_exception_flags |= float_flag_invalid;
> + return float16_default_nan(s);
> + }
> +
> + /* frec7(qNaN) = canonical NaN */
> + if (float16_is_quiet_nan(f, s)) {
> + return float16_default_nan(s);
> + }
> +
> + /* +-normal, +-subnormal */
> + uint64_t val = frec7(f, exp_size, frac_size, s);
> + return make_float16(val);
> +}
> +
> +static float32 frec7_s(float32 f, float_status *s)
> +{
> + int exp_size = 8, frac_size = 23;
> + bool sign = float32_is_neg(f);
> +
> + /* frec7(+-inf) = +-0 */
> + if (float32_is_infinity(f)) {
> + return float32_set_sign(float32_zero, sign);
> + }
> +
> + /* frec7(+-0) = +-inf */
> + if (float32_is_zero(f)) {
> + s->float_exception_flags |= float_flag_divbyzero;
> + return float32_set_sign(float32_infinity, sign);
> + }
> +
> + /* frec7(sNaN) = canonical NaN */
> + if (float32_is_signaling_nan(f, s)) {
> + s->float_exception_flags |= float_flag_invalid;
> + return float32_default_nan(s);
> + }
> +
> + /* frec7(qNaN) = canonical NaN */
> + if (float32_is_quiet_nan(f, s)) {
> + return float32_default_nan(s);
> + }
> +
> + /* +-normal, +-subnormal */
> + uint64_t val = frec7(f, exp_size, frac_size, s);
> + return make_float32(val);
> +}
> +
> +static float64 frec7_d(float64 f, float_status *s)
> +{
> + int exp_size = 11, frac_size = 52;
> + bool sign = float64_is_neg(f);
> +
> + /* frec7(+-inf) = +-0 */
> + if (float64_is_infinity(f)) {
> + return float64_set_sign(float64_zero, sign);
> + }
> +
> + /* frec7(+-0) = +-inf */
> + if (float64_is_zero(f)) {
> + s->float_exception_flags |= float_flag_divbyzero;
> + return float64_set_sign(float64_infinity, sign);
> + }
> +
> + /* frec7(sNaN) = canonical NaN */
> + if (float64_is_signaling_nan(f, s)) {
> + s->float_exception_flags |= float_flag_invalid;
> + return float64_default_nan(s);
> + }
> +
> + /* frec7(qNaN) = canonical NaN */
> + if (float64_is_quiet_nan(f, s)) {
> + return float64_default_nan(s);
> + }
> +
> + /* +-normal, +-subnormal */
> + uint64_t val = frec7(f, exp_size, frac_size, s);
> + return make_float64(val);
> +}
> +
> +RVVCALL(OPFVV1, vfrec7_v_h, OP_UU_H, H2, H2, frec7_h)
> +RVVCALL(OPFVV1, vfrec7_v_w, OP_UU_W, H4, H4, frec7_s)
> +RVVCALL(OPFVV1, vfrec7_v_d, OP_UU_D, H8, H8, frec7_d)
> +GEN_VEXT_V_ENV(vfrec7_v_h, 2, 2)
> +GEN_VEXT_V_ENV(vfrec7_v_w, 4, 4)
> +GEN_VEXT_V_ENV(vfrec7_v_d, 8, 8)
> +
> /* Vector Floating-Point MIN/MAX Instructions */
> RVVCALL(OPFVV2, vfmin_vv_h, OP_UUU_H, H2, H2, H2, float16_minimum_number)
> RVVCALL(OPFVV2, vfmin_vv_w, OP_UUU_W, H4, H4, H4, float32_minimum_number)
> --
> 2.25.1
>
>
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- Re: [PATCH v10 70/77] target/riscv: rvv-1.0: floating-point reciprocal estimate instruction,
Alistair Francis <=