fix edge case when chunk len is 1 and multithread worker num

encoding/kzg/prover/gpu/multiframe_proof.go

@@ -77,6 +77,8 @@ func (p *GpuComputeDevice) ComputeMultiFrameProof(polyFr []fr.Element, numChunks
 	l := chunkLen
 	numPoly := uint64(len(polyFr)) / dimE / chunkLen
 	fmt.Println("numPoly", numPoly)
+	fmt.Println("dimE", dimE)
+	fmt.Println("l", l)
 
 	begin := time.Now()
 
@@ -113,6 +115,16 @@ func (p *GpuComputeDevice) ComputeMultiFrameProof(polyFr []fr.Element, numChunks
 	}
 	preprocessDone := time.Now()
 
+	/*
+		for i := 0; i < int(l*numPoly); i++ {
+			vec := coeffStore[i]
+			for j := 0; j < len(vec); j++ {
+				fmt.Printf("%v ", vec[j].String())
+			}
+			fmt.Println()
+		}
+	*/
+
 	// Start using GPU
 	p.GpuLock.Lock()
 	defer p.GpuLock.Unlock()
@@ -125,18 +137,6 @@ func (p *GpuComputeDevice) ComputeMultiFrameProof(polyFr []fr.Element, numChunks
 	}
 	nttDone := time.Now()
 
-	/*
-		fmt.Println("after fft")
-		vec := gpu_utils.ConvertScalarFieldsToFrBytes(coeffStoreFft)
-		for i := 0; i < int(l*numPoly); i++ {
-			length := int(dimE) * 2
-			for j := 0; j < length; j++ {
-				fmt.Printf("%v ", vec[i*length+j].String())
-			}
-			fmt.Println()
-		}
-	*/
-
 	// transpose the FFT tranformed matrix
 	coeffStoreFftTranspose, err := Transpose(coeffStoreFft, int(l), int(numPoly), int(dimE)*2)
 	if err != nil {

encoding/test/main.go

@@ -52,7 +52,7 @@ func readpoints() {
 func TestKzgRs() {
 	isSmallTest := false
 
-	numSymbols := 4096 * 8
+	numSymbols := 4096 / 8
 	// encode parameters
 	numNode := uint64(4096) // 200
 	numSys := uint64(512)   // 180
@@ -72,7 +72,7 @@ func TestKzgRs() {
 	}
 
 	if isSmallTest {
-		numSymbols = 4
+		numSymbols = 2
 		numNode = 4
 		numSys = 2
 		numPar = numNode - numSys
@@ -96,7 +96,7 @@ func TestKzgRs() {
 	// create encoding object
 	p, _ := prover.NewProver(kzgConfig, true)
 
-	p.UseGpu = true
+	p.UseGpu = false
 
 	params := encoding.EncodingParams{NumChunks: numNode, ChunkLength: uint64(numSymbols) / uint64(numSys)}
 	enc, _ := p.GetKzgEncoder(params)
@@ -139,6 +139,11 @@ func TestKzgRs() {
 			log.Fatal("leading coset inconsistency")
 		}
 
+		// special case when chunk Len =1, we need to artificially use root doubled
+		if params.ChunkLength == uint64(1) {
+			j *= 2
+		}
+
 		lc := enc.Fs.ExpandedRootsOfUnity[uint64(j)]
 
 		g2Atn, err := kzg.ReadG2Point(uint64(len(f.Coeffs)), kzgConfig)

encoding/utils/gpu_utils/utils.go

@@ -85,6 +85,10 @@ func IcicleProjectiveToGnarkAffine(p bn254_icicle.Projective) bn254.G1Affine {
 func HostSliceIcicleProjectiveToGnarkAffine(ps core.HostSlice[bn254_icicle.Projective], numWorker int) []bn254.G1Affine {
 	output := make([]bn254.G1Affine, len(ps))
 
+	if len(ps) < numWorker {
+		numWorker = len(ps)
+	}
+
 	var wg sync.WaitGroup
 
 	interval := int(math.Ceil(float64(len(ps)) / float64(numWorker)))
@@ -112,6 +116,10 @@ func HostSliceIcicleProjectiveToGnarkAffine(ps core.HostSlice[bn254_icicle.Proje
 func ConvertFrToScalarFieldsBytesThread(data []fr.Element, numWorker int) []bn254_icicle.ScalarField {
 	scalars := make([]bn254_icicle.ScalarField, len(data))
 
+	if len(data) < numWorker {
+		numWorker = len(data)
+	}
+
 	var wg sync.WaitGroup
 
 	interval := int(math.Ceil(float64(len(data)) / float64(numWorker)))