diff mbox series

[71/76] target/arm: Implement increased precision FRSQRTE

Message ID 20250124162836.2332150-72-peter.maydell@linaro.org
State New
Headers show
Series target/arm: Implement FEAT_AFP and FEAT_RPRES | expand

Commit Message

Peter Maydell Jan. 24, 2025, 4:28 p.m. UTC 
Implement the increased precision variation of FRSQRTE.  In the
pseudocode this corresponds to the handling of the
"increasedprecision" boolean in the FPRSqrtEstimate() and
RecipSqrtEstimate() functions.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
 target/arm/vfp_helper.c | 77 ++++++++++++++++++++++++++++++++++-------
 1 file changed, 64 insertions(+), 13 deletions(-)
diff mbox series

Patch

diff --git a/target/arm/vfp_helper.c b/target/arm/vfp_helper.c
index 79e58c5bb2a..e63455c4bb9 100644
--- a/target/arm/vfp_helper.c
+++ b/target/arm/vfp_helper.c
@@ -1013,8 +1013,36 @@  static int do_recip_sqrt_estimate(int a)
     return estimate;
 }
 
+static int do_recip_sqrt_estimate_incprec(int a)
+{
+    /*
+     * The Arm ARM describes the 12-bit precision version of RecipSqrtEstimate
+     * in terms of an infinite-precision floating point calculation of a
+     * square root. We implement this using the same kind of pure integer
+     * algorithm as the 8-bit mantissa, to get the same bit-for-bit result.
+     */
+    int64_t b, estimate;
 
-static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
+    assert(1024 <= a && a < 4096);
+    if (a < 2048) {
+        a = a * 2 + 1;
+    } else {
+        a = (a >> 1) << 1;
+        a = (a + 1) * 2;
+    }
+    b = 8192;
+    while (a * (b + 1) * (b + 1) < (1ULL << 39)) {
+        b += 1;
+    }
+    estimate = (b + 1) / 2;
+
+    assert(4096 <= estimate && estimate < 8192);
+
+    return estimate;
+}
+
+static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac,
+                                    bool increasedprecision)
 {
     int estimate;
     uint32_t scaled;
@@ -1027,17 +1055,32 @@  static uint64_t recip_sqrt_estimate(int *exp , int exp_off, uint64_t frac)
         frac = extract64(frac, 0, 51) << 1;
     }
 
-    if (*exp & 1) {
-        /* scaled = UInt('01':fraction<51:45>) */
-        scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
+    if (increasedprecision) {
+        if (*exp & 1) {
+            /* scaled = UInt('01':fraction<51:42>) */
+            scaled = deposit32(1 << 10, 0, 10, extract64(frac, 42, 10));
+        } else {
+            /* scaled = UInt('1':fraction<51:41>) */
+            scaled = deposit32(1 << 11, 0, 11, extract64(frac, 41, 11));
+        }
+        estimate = do_recip_sqrt_estimate_incprec(scaled);
     } else {
-        /* scaled = UInt('1':fraction<51:44>) */
-        scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
+        if (*exp & 1) {
+            /* scaled = UInt('01':fraction<51:45>) */
+            scaled = deposit32(1 << 7, 0, 7, extract64(frac, 45, 7));
+        } else {
+            /* scaled = UInt('1':fraction<51:44>) */
+            scaled = deposit32(1 << 8, 0, 8, extract64(frac, 44, 8));
+        }
+        estimate = do_recip_sqrt_estimate(scaled);
     }
-    estimate = do_recip_sqrt_estimate(scaled);
 
     *exp = (exp_off - *exp) / 2;
-    return extract64(estimate, 0, 8) << 44;
+    if (increasedprecision) {
+        return extract64(estimate, 0, 12) << 40;
+    } else {
+        return extract64(estimate, 0, 8) << 44;
+    }
 }
 
 uint32_t HELPER(rsqrte_f16)(uint32_t input, float_status *s)
@@ -1076,7 +1119,7 @@  uint32_t HELPER(rsqrte_f16)(uint32_t input, float_status *s)
 
     f64_frac = ((uint64_t) f16_frac) << (52 - 10);
 
-    f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac);
+    f64_frac = recip_sqrt_estimate(&f16_exp, 44, f64_frac, false);
 
     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(2) */
     val = deposit32(0, 15, 1, f16_sign);
@@ -1125,12 +1168,20 @@  static float32 do_rsqrte_f32(float32 input, float_status *s, bool rpres)
 
     f64_frac = ((uint64_t) f32_frac) << 29;
 
-    f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac);
+    f64_frac = recip_sqrt_estimate(&f32_exp, 380, f64_frac, rpres);
 
-    /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(15) */
+    /*
+     * result = sign : result_exp<7:0> : estimate<7:0> : Zeros(15)
+     * or for increased precision
+     * result = sign : result_exp<7:0> : estimate<11:0> : Zeros(11)
+     */
     val = deposit32(0, 31, 1, f32_sign);
     val = deposit32(val, 23, 8, f32_exp);
-    val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
+    if (rpres) {
+        val = deposit32(val, 11, 12, extract64(f64_frac, 52 - 12, 12));
+    } else {
+        val = deposit32(val, 15, 8, extract64(f64_frac, 52 - 8, 8));
+    }
     return make_float32(val);
 }
 
@@ -1174,7 +1225,7 @@  float64 HELPER(rsqrte_f64)(float64 input, float_status *s)
         return float64_zero;
     }
 
-    f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac);
+    f64_frac = recip_sqrt_estimate(&f64_exp, 3068, f64_frac, false);
 
     /* result = sign : result_exp<4:0> : estimate<7:0> : Zeros(44) */
     val = deposit64(0, 61, 1, f64_sign);