diff options
| author | Jason Ekstrand <jason.ekstrand@intel.com> | 2016-12-09 09:34:50 -0800 |
|---|---|---|
| committer | Emil Velikov <emil.l.velikov@gmail.com> | 2016-12-15 16:46:28 +0000 |
| commit | fb9f0a1197e10f9b1c727b5b2956f36827308ad1 (patch) | |
| tree | d2d27d5e8ed04b9be2ae7d0062e45a229ebaf140 | |
| parent | 41c688a6c31ac5b985a3318e082f78103f061977 (diff) | |
| download | external_mesa3d-fb9f0a1197e10f9b1c727b5b2956f36827308ad1.zip external_mesa3d-fb9f0a1197e10f9b1c727b5b2956f36827308ad1.tar.gz external_mesa3d-fb9f0a1197e10f9b1c727b5b2956f36827308ad1.tar.bz2 | |
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@whitecape.org>
Cc: "13.0" <mesa-dev@lists.freedesktop.org>
(cherry picked from commit da1c49171d0df185545cfbbd600e287f7c6160fa)
| -rw-r--r-- | src/compiler/spirv/vtn_glsl450.c | 23 |
1 files 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]), |