From 28b402a778ab810beba454973e393b64c0b28013 Mon Sep 17 00:00:00 2001
From: Anton Mitrokhin <anton.mitrokhin@phystech.edu>
Date: Tue, 3 Nov 2015 13:31:05 +0300
Subject: [PATCH] [AVX-512]: transcendentals: add exp() implementation, TODO:
 log() and pow()

---
 builtins/target-avx512-common.ll | 70 +++++++++++++++++++++++++++++++-
 1 file changed, 69 insertions(+), 1 deletion(-)

diff --git a/builtins/target-avx512-common.ll b/builtins/target-avx512-common.ll
index 317b3ed8..b3ad98d2 100644
--- a/builtins/target-avx512-common.ll
+++ b/builtins/target-avx512-common.ll
@@ -1220,5 +1220,73 @@ declare_nvptx()
 rsqrtd_decl()
 rcpd_decl()
 
-transcendetals_decl()
+;; Transcendentals
+
+;; exponent
+define(`F144', `0x3FF7154760000000') ;; F144 = log(2, e)
+define(`D144', `0x3FF71547652B82FE') ;; D144 = log(2, e)
+
+declare <16 x float> @llvm.x86.avx512.exp2.ps(<16 x float>, <16 x float>, i16, i32) nounwind readnone
+declare <8 x double> @llvm.x86.avx512.exp2.pd(<8 x double>, <8 x double>, i8, i32) nounwind readnone
+declare <16 x float> @llvm.x86.avx512.mask.mul.ps.512(<16 x float>, <16 x float>, <16 x float>, i16, i32)
+declare <8 x double> @llvm.x86.avx512.mask.mul.pd.512(<8 x double>, <8 x double>, <8 x double>, i8, i32)
+
+define float @__exp_uniform_float(float %a) nounwind readnone alwaysinline {
+  %res = call float @__stdlib_expf(float %a)
+  ret float %res;
+}
+
+define double @__exp_uniform_double(double %a) nounwind readnone alwaysinline {
+  %res = call double @__stdlib_exp(double %a)
+  ret double %res;
+}
+
+define <16 x float> @__exp_varying_float(<16 x float> %a) nounwind readnone alwaysinline {
+  %a0 = call <16 x float> @llvm.x86.avx512.mask.mul.ps.512(<16 x float> <float F144, float F144, float F144, float F144,
+        float F144, float F144, float F144, float F144, float F144, float F144, float F144, float F144,
+        float F144, float F144, float F144, float F144>, <16 x float> %a, <16 x float> zeroinitializer, i16 -1, i32 0)
+  %res = call <16 x float> @llvm.x86.avx512.exp2.ps(<16 x float> %a0, <16 x float> zeroinitializer, i16 -1, i32 8)
+  ret <16 x float> %res
+}
+
+define <16 x double> @__exp_varying_double(<16 x double> %a) nounwind readnone alwaysinline {
+  %alo = shufflevector <16 x double> %a, <16 x double> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %ahi = shufflevector <16 x double> %a, <16 x double> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> 
+  %alo0 = call <8 x double> @llvm.x86.avx512.mask.mul.pd.512(<8 x double> <double D144, double D144, double D144, 
+          double D144, double D144, double D144, double D144, double D144>, <8 x double> %alo, <8 x double> zeroinitializer, i8 -1, i32 0)
+  %ahi0 = call <8 x double> @llvm.x86.avx512.mask.mul.pd.512(<8 x double> <double D144, double D144, double D144, 
+          double D144, double D144, double D144, double D144, double D144>, <8 x double> %ahi, <8 x double> zeroinitializer, i8 -1, i32 0)
+  %res_lo = call <8 x double> @llvm.x86.avx512.exp2.pd(<8 x double> %alo0, <8 x double> zeroinitializer, i8 -1, i32 8)
+  %res_hi = call <8 x double> @llvm.x86.avx512.exp2.pd(<8 x double> %ahi0, <8 x double> zeroinitializer, i8 -1, i32 8)
+  %res = shufflevector <8 x double> %res_lo, <8 x double> %res_hi, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> 
+  ret <16 x double> %res
+}
+
+;; power
+define float @__pow_uniform_float(float %a, float %b) nounwind readnone alwaysinline {
+  %res = call float @__stdlib_powf(float %a, float %b)
+  ret float %res;
+}
+
+define double @__pow_uniform_double(double %a, double %b) nounwind readnone alwaysinline {
+  %res = call double @__stdlib_pow(double %a, double %b)
+  ret double %res;
+}
+
+declare <16 x float> @__pow_varying_float(<16 x float> %a, <16 x float> %b) nounwind readnone alwaysinline
+
+;;define <16 x float> @__pow_varying_float(<16 x float> %a, <16 x float> %b) nounwind readnone alwaysinline
+;;  ret <16 x float> %a
+;;}
+
+declare <16 x double> @__pow_varying_double(<16 x double> %a, <16 x double> %b) nounwind readnone alwaysinline
+
+
+;; log
+declare float @__log_uniform_float(float %a) nounwind readnone alwaysinline
+declare double @__log_uniform_double(double %a) nounwind readnone alwaysinline
+declare <16 x float> @__log_varying_float(<16 x float> %a) nounwind readnone alwaysinline
+declare <16 x double> @__log_varying_double(<16 x double> %a) nounwind readnone alwaysinline
+
+;; Trigonometry
 trigonometry_decl()