From 9f9432d56c055b9704a76cad44da88d5e12f825c Mon Sep 17 00:00:00 2001 From: "Kristian H. Kristensen" Date: Thu, 27 Feb 2020 11:38:53 -0800 Subject: [PATCH] Revert "spirv: Use a simpler and more correct implementaiton of tanh()" This reverts commit da1c49171d0df185545cfbbd600e287f7c6160fa. The reduced formula has precision problems on fp16 around 0. Bring back the old formula, but make sure to keep the clamping. Tested-by: Marge Bot Part-of: --- src/compiler/spirv/vtn_glsl450.c | 27 +++++++++++++-------------- 1 file changed, 13 insertions(+), 14 deletions(-) diff --git a/src/compiler/spirv/vtn_glsl450.c b/src/compiler/spirv/vtn_glsl450.c index a074b6f3fb4..571f06cbed2 100644 --- a/src/compiler/spirv/vtn_glsl450.c +++ b/src/compiler/spirv/vtn_glsl450.c @@ -458,25 +458,24 @@ handle_glsl450_alu(struct vtn_builder *b, enum GLSLstd450 entrypoint, return; case GLSLstd450Tanh: { - /* tanh(x) := (0.5 * (e^x - e^(-x))) / (0.5 * (e^x + e^(-x))) + /* tanh(x) := (e^x - e^(-x)) / (e^x + e^(-x)) * - * With a little algebra this reduces to (e^2x - 1) / (e^2x + 1) + * We clamp x to [-10, +10] to avoid precision problems. When x > 10, + * e^x dominates the sum, e^(-x) is lost and tanh(x) is 1.0 for 32 bit + * floating point. * - * 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. - * - * For 16-bit precision we clamp x to (-inf, +4.2] since the maximum - * representable number is only 65,504 and e^(2*6) exceeds that. Also, - * if x > 4.2, tanh(x) will return 1.0 in fp16. + * For 16-bit precision this we clamp x to [-4.2, +4.2]. */ const uint32_t bit_size = src[0]->bit_size; const double clamped_x = bit_size > 16 ? 10.0 : 4.2; - nir_ssa_def *x = nir_fmin(nb, src[0], - nir_imm_floatN_t(nb, clamped_x, bit_size)); - nir_ssa_def *exp2x = build_exp(nb, nir_fmul_imm(nb, x, 2.0)); - val->ssa->def = nir_fdiv(nb, nir_fadd_imm(nb, exp2x, -1.0), - nir_fadd_imm(nb, exp2x, 1.0)); + nir_ssa_def *x = nir_fclamp(nb, src[0], + nir_imm_floatN_t(nb, -clamped_x, bit_size), + nir_imm_floatN_t(nb, clamped_x, bit_size)); + val->ssa->def = + nir_fdiv(nb, nir_fsub(nb, build_exp(nb, x), + build_exp(nb, nir_fneg(nb, x))), + nir_fadd(nb, build_exp(nb, x), + build_exp(nb, nir_fneg(nb, x)))); return; } -- 2.30.2