From c3fbd0dcb186c502390935b3088b711a24a6c7e4 Mon Sep 17 00:00:00 2001
From: Benjamin Lee <benjamin@computer.surgery>
Date: Mon, 22 Jan 2024 11:25:16 -0800
Subject: [PATCH] nak: use rro when emitting mufu on SM50

Fixes dEQP-VK.glsl.builtin.precision.*, which was previously failing for
trig functions, exp, and pow.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27203>
---
 src/nouveau/compiler/nak/builder.rs  | 47 ++++++++++++++++++++++++++++
 src/nouveau/compiler/nak/from_nir.rs | 14 ++-------
 2 files changed, 50 insertions(+), 11 deletions(-)

diff --git a/src/nouveau/compiler/nak/builder.rs b/src/nouveau/compiler/nak/builder.rs
index 67463ba4ee1..021e11a2d3c 100644
--- a/src/nouveau/compiler/nak/builder.rs
+++ b/src/nouveau/compiler/nak/builder.rs
@@ -456,6 +456,53 @@ pub trait SSABuilder: Builder {
         dst
     }
 
+    fn fsin(&mut self, src: Src) -> SSARef {
+        let tmp = if self.sm() >= 70 {
+            let frac_1_2pi = 1.0 / (2.0 * std::f32::consts::PI);
+            self.fmul(src, frac_1_2pi.into())
+        } else {
+            let tmp = self.alloc_ssa(RegFile::GPR, 1);
+            self.push_op(OpRro {
+                dst: tmp.into(),
+                op: RroOp::SinCos,
+                src,
+            });
+            tmp
+        };
+        self.mufu(MuFuOp::Sin, tmp.into())
+    }
+
+    fn fcos(&mut self, src: Src) -> SSARef {
+        let tmp = if self.sm() >= 70 {
+            let frac_1_2pi = 1.0 / (2.0 * std::f32::consts::PI);
+            self.fmul(src, frac_1_2pi.into())
+        } else {
+            let tmp = self.alloc_ssa(RegFile::GPR, 1);
+            self.push_op(OpRro {
+                dst: tmp.into(),
+                op: RroOp::SinCos,
+                src,
+            });
+            tmp
+        };
+        self.mufu(MuFuOp::Cos, tmp.into())
+    }
+
+    fn fexp2(&mut self, src: Src) -> SSARef {
+        let tmp = if self.sm() >= 70 {
+            src
+        } else {
+            let tmp = self.alloc_ssa(RegFile::GPR, 1);
+            self.push_op(OpRro {
+                dst: tmp.into(),
+                op: RroOp::Exp2,
+                src,
+            });
+            tmp.into()
+        };
+        self.mufu(MuFuOp::Exp2, tmp)
+    }
+
     fn prmt(&mut self, x: Src, y: Src, sel: [u8; 4]) -> SSARef {
         let dst = self.alloc_ssa(RegFile::GPR, 1);
         self.prmt_to(dst.into(), x, y, sel);
diff --git a/src/nouveau/compiler/nak/from_nir.rs b/src/nouveau/compiler/nak/from_nir.rs
index e1846806ef0..c1bb0f20260 100644
--- a/src/nouveau/compiler/nak/from_nir.rs
+++ b/src/nouveau/compiler/nak/from_nir.rs
@@ -714,11 +714,7 @@ impl<'a> ShaderFromNir<'a> {
                 }
                 dst
             }
-            nir_op_fcos => {
-                let frac_1_2pi = 1.0 / (2.0 * std::f32::consts::PI);
-                let tmp = b.fmul(srcs[0], frac_1_2pi.into());
-                b.mufu(MuFuOp::Cos, tmp.into())
-            }
+            nir_op_fcos => b.fcos(srcs[0]),
             nir_op_feq | nir_op_fge | nir_op_flt | nir_op_fneu => {
                 let src_type =
                     FloatType::from_bits(alu.get_src(0).bit_size().into());
@@ -756,7 +752,7 @@ impl<'a> ShaderFromNir<'a> {
                 }
                 dst
             }
-            nir_op_fexp2 => b.mufu(MuFuOp::Exp2, srcs[0]),
+            nir_op_fexp2 => b.fexp2(srcs[0]),
             nir_op_ffma => {
                 let ftype = FloatType::from_bits(alu.def.bit_size().into());
                 let dst;
@@ -938,11 +934,7 @@ impl<'a> ShaderFromNir<'a> {
                     panic!("Unsupported float type: f{}", alu.def.bit_size());
                 }
             }
-            nir_op_fsin => {
-                let frac_1_2pi = 1.0 / (2.0 * std::f32::consts::PI);
-                let tmp = b.fmul(srcs[0], frac_1_2pi.into());
-                b.mufu(MuFuOp::Sin, tmp.into())
-            }
+            nir_op_fsin => b.fsin(srcs[0]),
             nir_op_fsqrt => b.mufu(MuFuOp::Sqrt, srcs[0]),
             nir_op_i2f16 | nir_op_i2f32 | nir_op_i2f64 => {
                 let src_bits = alu.get_src(0).src.bit_size();