Use GPUToolbox's Float64-free math overrides on devices without fp64

[maleadt]

Several Base methods compute single-precision results in double precision, so they fail to compile for devices without Float64 support (e.g. Intel Arc): div, fld and cld of Float32 values on Julia 1.12 and 1.13, sind and friends, inversion and division of ComplexF32, sincospi/cispi, and comparisons between Float32/Float16 values and 32-bit integers. On such devices, this stacks GPUToolbox.Overlays.float64_overrides (JuliaGPU/GPUToolbox.jl#21, GPUToolbox 3.1) underneath SPIRVIntrinsics' method table, which replaces them with Float64-free implementations. Devices that support Float64 keep using Base's more accurate implementations.

This also removes the div(::Float32, ::Float32) = trunc(x / y) quirk, which truncated the rounded quotient and thus returned incorrect results on all devices, e.g. div(1f0, 0.1f0) == 10 instead of 9. Devices with Float64 now use Base's implementation, and devices without use the override above, which is exact whenever the quotient is.

I couldn't run this on hardware; on macOS, I checked that kernels calling div, fld, sind, inv// of ComplexF32 and hypot compile without any Float64 when targeting a device without Float64 support, and that they keep Base's implementations otherwise. The same stacking is used by OpenCL.jl, where the tests compile these functions for a device without Float64.

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic main) to apply these changes.

Click here to view the suggested changes.

diff --git a/src/compiler/compilation.jl b/src/compiler/compilation.jl
index f147240..7e36726 100644
--- a/src/compiler/compilation.jl
+++ b/src/compiler/compilation.jl
@@ -37,10 +37,12 @@ function GPUCompiler.method_table_view(job::oneAPICompilerJob)
         parent = SPIRVIntrinsics.method_table
     else
         # keep single-precision math that Base computes in Float64 out of double precision
-        parent = GPUCompiler.StackedMethodTable(job.world, SPIRVIntrinsics.method_table,
-                                                GPUToolbox.Overlays.float64_overrides)
+        parent = GPUCompiler.StackedMethodTable(
+            job.world, SPIRVIntrinsics.method_table,
+            GPUToolbox.Overlays.float64_overrides
+        )
     end
-    GPUCompiler.StackedMethodTable(job.world, method_table, parent)
+    return GPUCompiler.StackedMethodTable(job.world, method_table, parent)
 end
 
 # filter out OpenCL built-ins

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.66%. Comparing base (d2e06ce) to head (7cff0f8).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #640      +/-   ##
==========================================
+ Coverage   81.53%   81.66%   +0.13%     
==========================================
  Files          50       50              
  Lines        3607     3611       +4     
==========================================
+ Hits         2941     2949       +8     
+ Misses        666      662       -4     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.