Merge pull request #2030 from JuliaGPU/tb/fastmath

Add support for @cuda fastmath

Author: maleadt

Manifest.toml

@@ -142,9 +142,9 @@ version = "0.1.5"
 
 [[GPUCompiler]]
 deps = ["ExprTools", "InteractiveUtils", "LLVM", "Libdl", "Logging", "Scratch", "TimerOutputs", "UUIDs"]
-git-tree-sha1 = "72b2e3c2ba583d1a7aa35129e56cf92e07c083e3"
+git-tree-sha1 = "8de395b1243771bbb79ac832ec96c7def7a4586f"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
-version = "0.21.4"
+version = "0.22.0"
 
 [[InlineStrings]]
 deps = ["Parsers"]

Project.toml

@@ -46,7 +46,7 @@ Crayons = "4"
 DataFrames = "1"
 ExprTools = "0.1"
 GPUArrays = "8.6"
-GPUCompiler = "0.21"
+GPUCompiler = "0.22"
 KernelAbstractions = "0.9.2"
 LLVM = "6"
 Preferences = "1"

src/compiler/execution.jl

@@ -6,7 +6,7 @@ export @cuda, cudaconvert, cufunction, dynamic_cufunction, nextwarp, prevwarp
 ## high-level @cuda interface
 
 const MACRO_KWARGS = [:dynamic, :launch]
-const COMPILER_KWARGS = [:kernel, :name, :always_inline, :minthreads, :maxthreads, :blocks_per_sm, :maxregs]
+const COMPILER_KWARGS = [:kernel, :name, :always_inline, :minthreads, :maxthreads, :blocks_per_sm, :maxregs, :fastmath]
 const LAUNCH_KWARGS = [:cooperative, :blocks, :threads, :shmem, :stream]
 
 
@@ -306,6 +306,7 @@ The following keyword arguments are supported:
   supported on LLVM 4.0+)
 - `name`: override the name that the kernel will have in the generated code
 - `always_inline`: inline all function calls in the kernel
+- `fastmath`: use less precise square roots and flush denormals
 
 The output of this function is automatically cached, i.e. you can simply call `cufunction`
 in a hot path without degrading performance. New code will be generated automatically, when

test/core/codegen.jl

@@ -157,6 +157,32 @@ end
     @test !occursin(".local", asm)
 end
 
+@testset "fastmath" begin
+    function div_kernel(x)
+        i = threadIdx().x
+        @fastmath @inbounds x[i] = 1 / x[i]
+        return
+    end
+
+    asm = sprint(io->CUDA.code_ptx(io, div_kernel, Tuple{CuDeviceArray{Float32,1,AS.Global}}; fastmath=true))
+    @test occursin("div.approx.ftz", asm)
+
+    # libdevice only contains fast math versions of sqrt for CUDA 11.1+
+    if CUDA.runtime_version() >= v"11.1"
+        function sqrt_kernel(x)
+            i = threadIdx().x
+            @inbounds x[i] = sqrt(x[i])
+            return
+        end
+
+        asm = sprint(io->CUDA.code_ptx(io, sqrt_kernel, Tuple{CuDeviceArray{Float32,1,AS.Global}}))
+        @test occursin("sqrt.r", asm)
+
+        asm = sprint(io->CUDA.code_ptx(io, sqrt_kernel, Tuple{CuDeviceArray{Float32,1,AS.Global}}; fastmath=true))
+        @test occursin("sqrt.approx.ftz", asm)
+    end
+end
+
 end
 
 ############################################################################################