perf: remove allocs from iterator in sis for 16-512 params

field/generator/generator_sis.go

@@ -32,12 +32,14 @@ func generateSIS(F *config.Field, outputDir string) error {
 		FF               string
 		FieldPackagePath string
 		HasUnrolledFFT   bool
+		F31              bool
 	}
 
 	data := &sisTemplateData{
 		FF:               F.PackageName,
 		FieldPackagePath: fieldImportPath,
 		HasUnrolledFFT:   F.NbBytes == 32,
+		F31:              F.F31,
 	}
 
 	funcs := make(map[string]interface{})

field/generator/internal/templates/sis/sis.go.tmpl

@@ -169,11 +169,65 @@ func (r *RSis) Hash(v, res []{{ .FF }}.Element) error {
 	}
 	{{- end}}
 
-	// inner hash
-	it := NewLimbIterator(&VectorIterator{v: v}, r.LogTwoBound/8)
-	for i := 0; i < len(r.Ag); i++ {
-		r.InnerHash(it, res, k, r.kz, i, mask)
-	}
+	{{- if .F31}}
+		if r.Degree == 512 && r.LogTwoBound == 16 {
+			// this is our hot path, we don't use the iterator because with
+			// avx512 instructions, it actually ends up being most of the CPU time.
+			er := {{ .FF }}.Element{1} // mul by 1 --> mont reduce
+			polId := 0
+			var k512 [512]{{ .FF }}.Element
+			vk := {{ .FF }}.Vector(k512[:])
+			vRes := {{ .FF }}.Vector(res)
+			vb := {{ .FF }}.Vector(k512[256:]) 
+
+			cosets, err := r.Domain.CosetTable()
+			if err != nil {
+				return err
+			}
+			vCosets := {{ .FF }}.Vector(cosets)
+
+			for j := 0; j < len(v); j+=256 {
+				start := j 
+				end := j + 256
+				end = min(end, len(v))
+
+				// use half of vk to copy the v input to batch convert to regular form
+				copy(vb[:], v[start:end])
+				for k:= (end-start); k < 256; k++ {
+					vb[k][0] = 0
+				}
+				// batch montgomery -> regular
+				vb.ScalarMul(vb, &er)
+
+				// do the limb split 
+				for k := 0; k < 256; k++ {
+					k512[k*2][0] = uint32(uint16(vb[k][0]))
+					k512[k*2+1][0] = uint32(uint16(vb[k][0] >> 16))
+				}
+
+				// inner hash
+				vk.Mul(vk, vCosets)
+				r.Domain.FFT(k512[:], fft.DIF, fft.WithNbTasks(1))
+				vk.Mul(vk, {{.FF}}.Vector(r.Ag[polId]))
+				vRes.Add(vRes, vk)
+				polId++
+			}
+		} else {
+			// inner hash
+			it := NewLimbIterator(&VectorIterator{v: v}, r.LogTwoBound/8)
+			for i := 0; i < len(r.Ag); i++ {
+				r.InnerHash(it, res, k, r.kz, i, mask)
+			}
+		}
+	{{- else}}
+		// inner hash
+		it := NewLimbIterator(&VectorIterator{v: v}, r.LogTwoBound/8)
+		for i := 0; i < len(r.Ag); i++ {
+			r.InnerHash(it, res, k, r.kz, i, mask)
+		}
+	{{- end}}
+
+
 
 	// reduces mod Xᵈ+1
 	r.Domain.FFTInverse(res, fft.DIT, fft.OnCoset(), fft.WithNbTasks(1))