Mesa (master): spirv: Use a simpler and more correct implementaiton of tanh()

[Jason Ekstrand]

Module: Mesa
Branch: master
Commit: da1c49171d0df185545cfbbd600e287f7c6160fa
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=da1c49171d0df185545cfbbd600e287f7c6160fa

Author: Jason Ekstrand <jason.ekstrand at intel.com>
Date:   Fri Dec  9 09:34:50 2016 -0800

spirv: Use a simpler and more correct implementaiton of tanh()

The new implementation is more correct because it clamps the incoming value
to 10 to avoid floating-point overflow.  It also uses a much reduced
version of the formula which only requires 1 exp() rather than 2.  This
fixes all of the dEQP-VK.glsl.builtin.precision.tanh.* tests.

Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
Cc: "13.0" <mesa-dev at lists.freedesktop.org>

---

 src/compiler/spirv/vtn_glsl450.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)

diff --git a/src/compiler/spirv/vtn_glsl450.c b/src/compiler/spirv/vtn_glsl450.c
index cb0570d..fbc7ce6 100644
--- a/src/compiler/spirv/vtn_glsl450.c
+++ b/src/compiler/spirv/vtn_glsl450.c
@@ -565,16 +565,21 @@ handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint,
                                    build_exp(nb, nir_fneg(nb, src[0]))));
       return;
 
-   case GLSLstd450Tanh:
-      /* (0.5 * (e^x - e^(-x))) / (0.5 * (e^x + e^(-x))) */
-      val->ssa->def =
-         nir_fdiv(nb, nir_fmul(nb, nir_imm_float(nb, 0.5f),
-                                   nir_fsub(nb, build_exp(nb, src[0]),
-                                                build_exp(nb, nir_fneg(nb, src[0])))),
-                      nir_fmul(nb, nir_imm_float(nb, 0.5f),
-                                   nir_fadd(nb, build_exp(nb, src[0]),
-                                                build_exp(nb, nir_fneg(nb, src[0])))));
+   case GLSLstd450Tanh: {
+      /* tanh(x) := (0.5 * (e^x - e^(-x))) / (0.5 * (e^x + e^(-x)))
+       *
+       * With a little algebra this reduces to (e^2x - 1) / (e^2x + 1)
+       *
+       * We clamp x to (-inf, +10] to avoid precision problems.  When x > 10,
+       * e^2x is so much larger than 1.0 that 1.0 gets flushed to zero in the
+       * computation e^2x +/- 1 so it can be ignored.
+       */
+      nir_ssa_def *x = nir_fmin(nb, src[0], nir_imm_float(nb, 10));
+      nir_ssa_def *exp2x = build_exp(nb, nir_fmul(nb, x, nir_imm_float(nb, 2)));
+      val->ssa->def = nir_fdiv(nb, nir_fsub(nb, exp2x, nir_imm_float(nb, 1)),
+                                   nir_fadd(nb, exp2x, nir_imm_float(nb, 1)));
       return;
+   }
 
    case GLSLstd450Asinh:
       val->ssa->def = nir_fmul(nb, nir_fsign(nb, src[0]),