kernels: import go-simd/floats for sum, elementwise, extrema, clip, exp, log (K1)

README.md

@@ -175,6 +175,19 @@ arch is therefore a durable foundation. The numeric loops live in
 on every 64-bit target Go supports (amd64, arm64, ppc64le, loong64, riscv64,
 s390x), behind the same API and held to the same tests.
 
+The one-dimensional float64 kernels that other libraries need too live once, in
+[go-simd/floats](https://github.com/go-simd/floats) (K1 of the go-scikit
+[performance campaign](https://github.com/go-scikit/roadmap/blob/main/docs/performance-campaign.md)),
+which this module imports: on amd64 (AVX2) and arm64 (NEON) the elementwise
+`Add`/`Sub`/`Mul`/`Div`, `Sqrt`, `Max`/`Min` and `Clip` kernels, and everywhere
+`Exp`, `Log` and the correctly rounded float32 `Exp`/`Log` (the AVX2 and NEON
+kernels were written here and moved down). **`Sum` is NumPy's pairwise sum**
+(`floats.SumPairwise`): bit-identical to `np.sum` of a contiguous 1-D float64
+array, on every architecture, at every size. It replaces lane-parallel sums whose
+last bits differed from NumPy's and from one architecture to the next, and it is
+serial (a multicore split would regroup NumPy's order). The other architectures
+keep their own elementwise kernels here until floats has them.
+
 ## Goroutines
 
 Operations on large arrays are spread over `GOMAXPROCS` goroutines. After the

go.mod

@@ -2,4 +2,9 @@ module github.com/go-ndarray/ndarray
 
 go 1.27.2
 
-require github.com/go-fft/fft v0.28.11
+require (
+	github.com/go-fft/fft v0.28.11
+	github.com/go-simd/floats v0.9.0
+)
+
+require golang.org/x/sys v0.49.0 // indirect

go.sum

@@ -1,2 +1,6 @@
 github.com/go-fft/fft v0.28.11 h1:ZfT78fSmPjZU2E9PgkSI/0f0E3sesYpl13GHmIFzRGc=
 github.com/go-fft/fft v0.28.11/go.mod h1:GN5V4TWCjV05ECqH1ONfjFM6Z1m2shqCdssxKGqQ41s=
+github.com/go-simd/floats v0.9.0 h1:P/sMlwYWwaqCah/cMcBRfffK1JcUxXZVj76f4AwQRN0=
+github.com/go-simd/floats v0.9.0/go.mod h1:Aq5a6oY5pUqm9Hpi+iV2nrNmaaUb0rwHvUHIerXWSvA=
+golang.org/x/sys v0.49.0 h1:XbzkgYJHdqh/8m2Uu0W/dQv8nktxx4BFHp1M0gROTXA=
+golang.org/x/sys v0.49.0/go.mod h1:hNLxWAXmnKAxqDtdwIYC4bM9oQPEecfsnNMuSxOs3og=

internal/kernels/asmgen/amd64/gen.go

@@ -37,14 +37,11 @@ import (
 
 func main() {
 	f := emit.NewFile("amd64")
-	f.Add(sumKernel())
-	f.Add(sqrtKernel())
-	f.Add(extremeKernel("maxSSE2", "MAXPD", "MAXSD"))
-	f.Add(extremeKernel("minSSE2", "MINPD", "MINSD"))
-	f.Add(binKernel("addSSE2", "ADDPD", "ADDSD"))
-	f.Add(binKernel("subSSE2", "SUBPD", "SUBSD"))
-	f.Add(binKernel("mulSSE2", "MULPD", "MULSD"))
-	f.Add(binKernel("divSSE2", "DIVPD", "DIVSD"))
+	// The float64 sum, sqrt, max/min and add/sub/mul/div kernels, and the float32
+	// exp/log and float64 log kernels, moved to go-simd/floats (K1 of the
+	// go-scikit performance campaign): sumKernel, sqrtKernel, extremeKernel,
+	// binKernel, exp32CRKernel, log32CRKernel and logKernel below are no longer
+	// emitted.
 	f.Add(gemmFMAKernel())
 	f.Add(gemm32FMAKernel())
 	for mr := 1; mr <= gemmMR; mr++ {
@@ -54,10 +51,6 @@ func main() {
 	f.Add(dotFMAKernel())
 	f.Add(gemmSSE2HalfKernel())
 	f.Add(amd64.FeatureProbe("hasFMA", amd64.FMA))
-	f.Add(amd64.FeatureProbe("hasAVX2", amd64.AVX2))
-	f.Add(exp32CRKernel())
-	f.Add(log32CRKernel())
-	f.Add(logKernel())
 	f.Add(blocks8Kernel())
 	f.Add(bin32Kernel("add32SSE2", "ADDPS", "ADDSS"))
 	f.Add(bin32Kernel("sub32SSE2", "SUBPS", "SUBSS"))

internal/kernels/asmgen/arm64/gen.go

@@ -41,20 +41,15 @@ func reduceSig() arm64.Signature {
 
 func main() {
 	f := emit.NewFile("arm64")
-	f.Add(sumKernel())
+	// The float64 sum, sqrt, add/sub/mul/div and max/min kernels moved to
+	// go-simd/floats (K1 of the go-scikit performance campaign): sumKernel,
+	// sqrtKernel, binKernel and extremumKernel below are no longer emitted.
 	f.Add(dotKernel())
-	f.Add(sqrtKernel())
-	f.Add(binKernel("addNEON", "VFADD", "FADDD"))
-	f.Add(binKernel("subNEON", "VFSUB", "FSUBD"))
-	f.Add(binKernel("mulNEON", "VFMUL", "FMULD"))
-	f.Add(binKernel("divNEON", "VFDIV", "FDIVD"))
 	f.Add(bin32Kernel("add32NEON", "VFADD", "FADDS"))
 	f.Add(bin32Kernel("sub32NEON", "VFSUB", "FSUBS"))
 	f.Add(bin32Kernel("mul32NEON", "VFMUL", "FMULS"))
 	f.Add(bin32Kernel("div32NEON", "VFDIV", "FDIVS"))
 	f.Add(sqrt32Kernel())
-	f.Add(extremumKernel("maxNEON", "VFMAX", "FMAXD"))
-	f.Add(extremumKernel("minNEON", "VFMIN", "FMIND"))
 	f.Add(extremum32Kernel("max32NEON", "VFMAX", "FMAXS"))
 	f.Add(extremum32Kernel("min32NEON", "VFMIN", "FMINS"))
 	f.Add(gemmKernel())