Implement the increased precision variation of FRECPE. In the
pseudocode this corresponds to the handling of the
"increasedprecision" boolean in the FPRecipEstimate() and
RecipEstimate() functions.
Signed-off-by: Peter Maydell <peter.mayd...@linaro.org>
---
target/arm/vfp_helper.c | 54 +++++++++++++++++++++++++++++++++++------
1 file changed, 46 insertions(+), 8 deletions(-)
diff --git a/target/arm/vfp_helper.c b/target/arm/vfp_helper.c
index 1b7ecc14621..79e58c5bb2a 100644
--- a/target/arm/vfp_helper.c
+++ b/target/arm/vfp_helper.c
@@ -731,6 +731,33 @@ static int recip_estimate(int input)
return r;
}
+/*
+ * Increased precision version:
+ * input is a 13 bit fixed point number
+ * input range 2048 .. 4095 for a number from 0.5 <= x < 1.0.
+ * result range 4096 .. 8191 for a number from 1.0 to 2.0
+ */
+static int recip_estimate_incprec(int input)
+{
+ int a, b, r;
+ assert(2048 <= input && input < 4096);
+ a = (input * 2) + 1;
+ /*
+ * The pseudocode expresses this as an operation on infinite
+ * precision reals where it calculates 2^25 / a and then looks
+ * at the error between that and the rounded-down-to-integer
+ * value to see if it should instead round up. We instead
+ * follow the same approach as the pseudocode for the 8-bit
+ * precision version, and calculate (2 * (2^25 / a)) as an
+ * integer so we can do the "add one and halve" to round it.
+ * So the 1 << 26 here is correct.
+ */
+ b = (1 << 26) / a;
+ r = (b + 1) >> 1;
+ assert(4096 <= r && r < 8192);
+ return r;
+}
+
/*
* Common wrapper to call recip_estimate
*
@@ -740,7 +767,8 @@ static int recip_estimate(int input)
* callee.
*/
-static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac)
+static uint64_t call_recip_estimate(int *exp, int exp_off, uint64_t frac,
+ bool increasedprecision)
{
uint32_t scaled, estimate;
uint64_t result_frac;
@@ -756,12 +784,22 @@ static uint64_t call_recip_estimate(int *exp, int
exp_off, uint64_t frac)
}
}
- /* scaled = UInt('1':fraction<51:44>) */
- scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
- estimate = recip_estimate(scaled);
+ if (increasedprecision) {
+ /* scaled = UInt('1':fraction<51:41>) */
+ scaled = deposit32(1 << 11, 0, 11, extract64(frac, 41, 11));
+ estimate = recip_estimate_incprec(scaled);
+ } else {
+ /* scaled = UInt('1':fraction<51:44>) */
+ scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
+ estimate = recip_estimate(scaled);
+ }
result_exp = exp_off - *exp;
- result_frac = deposit64(0, 44, 8, estimate);
+ if (increasedprecision) {
+ result_frac = deposit64(0, 40, 12, estimate);
+ } else {
+ result_frac = deposit64(0, 44, 8, estimate);
+ }
if (result_exp == 0) {
result_frac = deposit64(result_frac >> 1, 51, 1, 1);
} else if (result_exp == -1) {
@@ -830,7 +868,7 @@ uint32_t HELPER(recpe_f16)(uint32_t input, float_status
*fpst)
}
f64_frac = call_recip_estimate(&f16_exp, 29,
- ((uint64_t) f16_frac) << (52 - 10));
+ ((uint64_t) f16_frac) << (52 - 10), false);
/* result = sign : result_exp<4:0> : fraction<51:42> */
f16_val = deposit32(0, 15, 1, f16_sign);
@@ -883,7 +921,7 @@ static float32 do_recpe_f32(float32 input, float_status
*fpst, bool rpres)
}
f64_frac = call_recip_estimate(&f32_exp, 253,
- ((uint64_t) f32_frac) << (52 - 23));
+ ((uint64_t) f32_frac) << (52 - 23), rpres);
/* result = sign : result_exp<7:0> : fraction<51:29> */
f32_val = deposit32(0, 31, 1, f32_sign);
@@ -941,7 +979,7 @@ float64 HELPER(recpe_f64)(float64 input, float_status *fpst)
return float64_set_sign(float64_zero, float64_is_neg(f64));
}
- f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac);
+ f64_frac = call_recip_estimate(&f64_exp, 2045, f64_frac, false);
/* result = sign : result_exp<10:0> : fraction<51:0>; */
f64_val = deposit64(0, 63, 1, f64_sign);
--
2.34.1
