nir/flrp: Lower flrp(#a, #b, c) differently

If the magnitudes of #a and #b are such that (b-a) won't lose too much
precision, lower as a+c(b-a).

No changes on any other Intel platforms.

v2: Rebase on 424372e5dd5 ("nir: Use the flrp lowering pass instead of
nir_opt_algebraic")

Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 8192503 -> 8192383 (<.01%)
instructions in affected programs: 18417 -> 18297 (-0.65%)
helped: 68
HURT: 0
helped stats (abs) min: 1 max: 18 x̄: 1.76 x̃: 1
helped stats (rel) min: 0.19% max: 7.89% x̄: 1.10% x̃: 0.43%
95% mean confidence interval for instructions value: -2.48 -1.05
95% mean confidence interval for instructions %-change: -1.56% -0.63%
Instructions are helped.

total cycles in shared programs: 188662536 -> 188661956 (<.01%)
cycles in affected programs: 744476 -> 743896 (-0.08%)
helped: 62
HURT: 0
helped stats (abs) min: 4 max: 60 x̄: 9.35 x̃: 6
helped stats (rel) min: 0.02% max: 4.84% x̄: 0.27% x̃: 0.06%
95% mean confidence interval for cycles value: -12.37 -6.34
95% mean confidence interval for cycles %-change: -0.48% -0.06%
Cycles are helped.

Reviewed-by: Matt Turner <mattst88@gmail.com>

diff --git a/src/compiler/nir/nir_lower_flrp.c b/src/compiler/nir/nir_lower_flrp.c
index 2d57998b41d4b70a3687e87e3ca808df844e3c39..952068ec9cc4afbd64d5c3985547dc7dba0ddace 100644 (file)
--- a/src/compiler/nir/nir_lower_flrp.c
+++ b/src/compiler/nir/nir_lower_flrp.c
@@ -20,6 +20,7 @@
  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  * IN THE SOFTWARE.
  */
+#include <math.h>
 #include "nir.h"
 #include "nir_builder.h"
 #include "util/u_vector.h"
@@ -136,6 +137,58 @@ replace_with_fast(struct nir_builder *bld, struct u_vector *dead_flrp,
    append_flrp_to_dead_list(dead_flrp, alu);
 }
 
+static bool
+sources_are_constants_with_similar_magnitudes(const nir_alu_instr *instr)
+{
+   nir_const_value *val0 = nir_src_as_const_value(instr->src[0].src);
+   nir_const_value *val1 = nir_src_as_const_value(instr->src[1].src);
+
+   if (val0 == NULL || val1 == NULL)
+      return false;
+
+   const uint8_t *const swizzle0 = instr->src[0].swizzle;
+   const uint8_t *const swizzle1 = instr->src[1].swizzle;
+   const unsigned num_components = nir_dest_num_components(instr->dest.dest);
+
+   if (instr->dest.dest.ssa.bit_size == 32) {
+      for (unsigned i = 0; i < num_components; i++) {
+         int exp0;
+         int exp1;
+
+         frexpf(val0[swizzle0[i]].f32, &exp0);
+         frexpf(val1[swizzle1[i]].f32, &exp1);
+
+         /* If the difference between exponents is >= 24, then A+B will always
+          * have the value whichever between A and B has the largest absolute
+          * value.  So, [0, 23] is the valid range.  The smaller the limit
+          * value, the more precision will be maintained at a potential
+          * performance cost.  Somewhat arbitrarilly split the range in half.
+          */
+         if (abs(exp0 - exp1) > (23 / 2))
+            return false;
+      }
+   } else {
+      for (unsigned i = 0; i < num_components; i++) {
+         int exp0;
+         int exp1;
+
+         frexp(val0[swizzle0[i]].f64, &exp0);
+         frexp(val1[swizzle1[i]].f64, &exp1);
+
+         /* If the difference between exponents is >= 53, then A+B will always
+          * have the value whichever between A and B has the largest absolute
+          * value.  So, [0, 52] is the valid range.  The smaller the limit
+          * value, the more precision will be maintained at a potential
+          * performance cost.  Somewhat arbitrarilly split the range in half.
+          */
+         if (abs(exp0 - exp1) > (52 / 2))
+            return false;
+      }
+   }
+
+   return true;
+}
+
 static void
 convert_flrp_instruction(nir_builder *bld,
                          struct u_vector *dead_flrp,
@@ -197,6 +250,21 @@ convert_flrp_instruction(nir_builder *bld,
       return;
    }
 
+   /*
+    * - If x and y are both immediates and the relative magnitude of the
+    *   values is similar (such that x-y does not lose too much precision):
+    *
+    *        x + t(x - y)
+    *
+    *   We rely on constant folding to eliminate x-y, and we rely on
+    *   nir_opt_algebraic to possibly generate an FMA.  The cost is either one
+    *   FMA or two instructions.
+    */
+   if (sources_are_constants_with_similar_magnitudes(alu)) {
+      replace_with_fast(bld, dead_flrp, alu);
+      return;
+   }
+
    if (have_ffma) {
       if (always_precise) {
          replace_with_strict_ffma(bld, dead_flrp, alu);