Disable sqrt fast math test on CUDA 11.0.

JuliaGPU · Aug 17, 2023 · a2d3219 · a2d3219
1 parent b85130d
commit a2d3219
Showing 1 changed file with 15 additions and 12 deletions.
diff --git a/test/core/codegen.jl b/test/core/codegen.jl
@@ -158,26 +158,29 @@ end
 end
 
 @testset "fastmath" begin
-    function sqrt_kernel(x)
-        i = threadIdx().x
-        @inbounds x[i] = sqrt(x[i])
-        return
-    end
-
     function div_kernel(x)
         i = threadIdx().x
         @fastmath @inbounds x[i] = 1 / 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)
-
     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