sm9/bn256: ppc64x gfp #256

sm9/bn256/gfp_decl.go

@@ -1,4 +1,4 @@
-//go:build (amd64 || arm64) && !purego
+//go:build (amd64 || arm64 || ppc64 || ppc64le) && !purego
 
 package bn256
 

sm9/bn256/gfp_generic.go

@@ -1,4 +1,4 @@
-//go:build purego || !(amd64 || arm64)
+//go:build purego || !(amd64 || arm64 || ppc64 || ppc64le)
 
 package bn256
 

sm9/bn256/gfp_ppc64x.go

@@ -1,53 +0,0 @@
-// Copyright 2024 Sun Yimin. All rights reserved.
-// Use of this source code is governed by a MIT-style
-// license that can be found in the LICENSE file.
-
-//go:build (ppc64 || ppc64le) && !purego
-
-package bn256
-
-// Set c = p - a, if c == p, then c = 0
-// It seems this function's performance is worse than gfpSub with zero.
-//
-//go:noescape
-func gfpNegAsm(c, a *gfP)
-
-// Set c = a + b, if c >= p, then c = c - p
-//
-//go:noescape
-func gfpAddAsm(c, a, b *gfP)
-
-// Set c = a + a
-//
-//go:noescape
-func gfpDoubleAsm(c, a *gfP)
-
-// Set c = a + a + a
-//
-//go:noescape
-func gfpTripleAsm(c, a *gfP)
-
-// Set c = a - b, if c is negative, then c = c + p
-//
-//go:noescape
-func gfpSubAsm(c, a, b *gfP)
-
-// Montgomery multiplication. Sets res = in1 * in2 * R⁻¹ mod p.
-//
-//go:noescape
-func gfpMulAsm(c, a, b *gfP)
-
-// Montgomery square, repeated n times (n >= 1).
-//
-//go:noescape
-func gfpSqrAsm(res, in *gfP, n int)
-
-// Marshal gfP into big endian form
-//
-//go:noescape
-func gfpMarshalAsm(out *[32]byte, in *gfP)
-
-// Unmarshal the bytes into little endian form
-//
-//go:noescape
-func gfpUnmarshalAsm(out *gfP, in *[32]byte)

sm9/bn256/gfp_ppc64x.s

@@ -7,19 +7,20 @@
 #include "textflag.h"
 
 //func gfpUnmarshal(out *gfP, in *[32]byte)
-TEXT ·gfpUnmarshalAsm(SB), NOSPLIT, $0-16
+TEXT ·gfpUnmarshal(SB), NOSPLIT, $0-16
 	MOVD	res+0(FP), R3
 	MOVD	in+8(FP), R4
 	BR	gfpInternalEndianSwap<>(SB)
 
 // func gfpMarshal(out *[32]byte, in *gfP)
-TEXT ·gfpMarshalAsm(SB), NOSPLIT, $0-16
+TEXT ·gfpMarshal(SB), NOSPLIT, $0-16
 	MOVD	res+0(FP), R3
 	MOVD	in+8(FP), R4
 	BR	gfpInternalEndianSwap<>(SB)
 
 TEXT gfpInternalEndianSwap<>(SB), NOSPLIT, $0-0
 	// Index registers needed for BR movs
+#ifdef GOARCH_ppc64le	
 	MOVD	$8, R9
 	MOVD	$16, R10
 	MOVD	$24, R14
@@ -33,7 +34,17 @@ TEXT gfpInternalEndianSwap<>(SB), NOSPLIT, $0-0
 	MOVD	R7, 8(R3)
 	MOVD	R6, 16(R3)
 	MOVD	R5, 24(R3)
+#else
+	MOVD	$16, R10
+	LXVD2X (R4)(R0), V0
+	LXVD2X (R4)(R10), V1
 
+	XXPERMDI V0, V0, $2, V0
+	XXPERMDI V1, V1, $2, V1
+
+	STXVD2X V1, (R0+R3)
+	STXVD2X V0, (R10+R3)	
+#endif
 	RET
 
 #define X1L   V0
@@ -70,7 +81,7 @@ TEXT gfpInternalEndianSwap<>(SB), NOSPLIT, $0-0
 	VSEL     TT0, T0, SEL1, T0  \
 	VSEL     TT1, T1, SEL1, T1  \
 
-TEXT ·gfpNegAsm(SB),0,$0-16
+TEXT ·gfpNeg(SB),0,$0-16
 	MOVD c+0(FP), R3
 	MOVD a+8(FP), R4
 
@@ -98,7 +109,7 @@ TEXT ·gfpNegAsm(SB),0,$0-16
 	STXVD2X T1, (R5+R3)
 	RET
 
-TEXT ·gfpSubAsm(SB),0,$0-24
+TEXT ·gfpSub(SB),0,$0-24
 	MOVD c+0(FP), R3
 	MOVD a+8(FP), R4
 	MOVD b+16(FP), R5
@@ -144,7 +155,7 @@ TEXT ·gfpSubAsm(SB),0,$0-24
 	VSEL     TT0, T0, SEL1, T0    \
 	VSEL     TT1, T1, SEL1, T1
 
-TEXT ·gfpAddAsm(SB),0,$0-24
+TEXT ·gfpAdd(SB),0,$0-24
 	MOVD c+0(FP), R3
 	MOVD a+8(FP), R4
 	MOVD b+16(FP), R5
@@ -177,7 +188,7 @@ TEXT ·gfpAddAsm(SB),0,$0-24
 	STXVD2X T1, (R6+R3)
 	RET
 
-TEXT ·gfpDoubleAsm(SB),0,$0-16
+TEXT ·gfpDouble(SB),0,$0-16
 	MOVD c+0(FP), R3
 	MOVD a+8(FP), R4
 
@@ -204,7 +215,7 @@ TEXT ·gfpDoubleAsm(SB),0,$0-16
 	STXVD2X T1, (R6+R3)
 	RET
 
-TEXT ·gfpTripleAsm(SB),0,$0-16
+TEXT ·gfpTriple(SB),0,$0-16
 	MOVD c+0(FP), R3
 	MOVD a+8(FP), R4
 
@@ -726,7 +737,7 @@ TEXT gfpMulInternal<>(SB), NOSPLIT, $0
 #define T1    V7
 #define K0    V31
 
-TEXT ·gfpMulAsm(SB),NOSPLIT,$0
+TEXT ·gfpMul(SB),NOSPLIT,$0
 	MOVD	c+0(FP), res_ptr
 	MOVD	a+8(FP), x_ptr
 	MOVD	b+16(FP), y_ptr
@@ -766,7 +777,7 @@ TEXT ·gfpMulAsm(SB),NOSPLIT,$0
 	RET
 
 // func gfpSqr(res, in *gfP, n int)
-TEXT ·gfpSqrAsm(SB),NOSPLIT,$0
+TEXT ·gfpSqr(SB),NOSPLIT,$0
 	MOVD res+0(FP), res_ptr
 	MOVD in+8(FP), x_ptr
 	MOVD n+16(FP), N
@@ -825,3 +836,102 @@ done:
 #undef T0
 #undef T1
 #undef K0
+
+
+/* ---------------------------------------*/
+#define res_ptr R3
+#define x_ptr R4
+#define CPOOL R7
+
+#define M0    V5
+#define M1    V4
+#define T0    V6
+#define T1    V7
+#define T2    V8
+
+#define ADD1  V16
+#define ADD1H V17
+#define ADD2  V18
+#define ADD2H V19
+#define RED1  V20
+#define RED1H V21
+#define RED2  V22
+#define RED2H V23
+#define CAR1  V24
+#define CAR1M V25
+
+#define MK0   V30
+#define K0    V31
+
+// TMP1, TMP2 used in
+// VMULT macros
+#define TMP1  V13
+#define TMP2  V27
+#define ONE   V29 // 1s splatted by word
+// func gfpFromMont(res, in *gfP)
+TEXT ·gfpFromMont(SB),NOSPLIT,$0
+	MOVD res+0(FP), res_ptr
+	MOVD in+8(FP), x_ptr
+
+	MOVD $16, R16
+
+	LXVD2X (R0)(x_ptr), T0
+	LXVD2X (R16)(x_ptr), T1
+
+	XXPERMDI T0, T0, $2, T0
+	XXPERMDI T1, T1, $2, T1
+
+	MOVD $·p2+0(SB), CPOOL
+	LXVD2X (CPOOL)(R0), M0
+	LXVD2X (CPOOL)(R16), M1
+	
+	XXPERMDI M0, M0, $2, M0
+	XXPERMDI M1, M1, $2, M1
+
+	MOVD $·np+0(SB), CPOOL
+	LXVD2X (CPOOL)(R0), K0
+	VSPLTW $1, K0, K0
+
+	// ---------------------------------------------------------------------------/
+	VSPLTISW $1, ONE
+	VSPLTISB $0, T2 // VZERO T2
+
+	MOVD $8, R5
+	MOVD R5, CTR
+
+loop:
+	VMULUWM T0, K0, MK0
+	VSPLTW $3, MK0, MK0
+
+	VMULT_ADD(M0, MK0, T0, ONE, RED1, RED1H)
+	VMULT_ADD(M1, MK0, T1, ONE, RED2, RED2H)
+
+	VSLDOI $12, RED2, RED1, RED1 // VSLDB
+	VSLDOI $12, T2, RED2, RED2   // VSLDB
+
+	VADDCUQ RED1H, RED1, CAR1M   // VACCQ
+	VADDUQM RED1H, RED1, T0      // VAQ
+
+	// << ready for next MK0
+
+	VADDECUQ RED2H, RED2, CAR1M, T2    // VACCCQ
+	VADDEUQM RED2H, RED2, CAR1M, T1    // VACQ
+
+	BDNZ loop
+	// ---------------------------------------------------
+	VSPLTISB $0, RED1 // VZERO RED1
+	VSUBCUQ  T0, M0, CAR1         // VSCBIQ
+	VSUBUQM  T0, M0, ADD1         // VSQ
+	VSUBECUQ T1, M1, CAR1, CAR1M  // VSBCBIQ
+	VSUBEUQM T1, M1, CAR1, ADD2   // VSBIQ
+	VSUBEUQM T2, RED1, CAR1M, T2  // VSBIQ
+
+	// what output to use, ADD2||ADD1 or T1||T0?
+	VSEL ADD1, T0, T2, T0
+	VSEL ADD2, T1, T2, T1
+
+	XXPERMDI T0, T0, $2, T0
+	XXPERMDI T1, T1, $2, T1
+	STXVD2X T0, (R0)(res_ptr)
+	STXVD2X T1, (R16)(res_ptr)	
+	RET

sm9/bn256/gfp_ppc64x_test.go

@@ -1,138 +0,0 @@
-// Copyright 2024 Sun Yimin. All rights reserved.
-// Use of this source code is governed by a MIT-style
-// license that can be found in the LICENSE file.
-
-//go:build (ppc64 || ppc64le) && !purego
-
-package bn256
-
-import "testing"
-
-func TestGfpNegAsm(t *testing.T) {
-	x := fromBigInt(bigFromHex("9093a2b979e6186f43a9b28d41ba644d533377f2ede8c66b19774bf4a9c7a596"))
-	got := &gfP{}
-	gfpSubAsm(got, zero, x)
-	expected := &gfP{}
-	gfpNegAsm(expected, x)
-	if *expected != *got {
-		t.Errorf("got %v, expected %v", got, expected)
-	}
-	gfpSubAsm(got, zero, zero)
-	gfpNegAsm(expected, zero)
-	if *expected != *got {
-		t.Errorf("got %v, expected %v", got, expected)
-	}
-}
-
-func TestGfpAsmBasicOperations(t *testing.T) {
-	x := fromBigInt(bigFromHex("85AEF3D078640C98597B6027B441A01FF1DD2C190F5E93C454806C11D8806141"))
-	y := fromBigInt(bigFromHex("3722755292130B08D2AAB97FD34EC120EE265948D19C17ABF9B7213BAF82D65B"))
-	expectedAdd := fromBigInt(bigFromHex("0691692307d370af56226e57920199fbbe10f216c67fbc9468c7f225a4b1f21f"))
-	expectedDouble := fromBigInt(bigFromHex("551de7a0ee24723edcf314ff72f478fac1c7c4e7044238acc3913cfbcdaf7d05"))
-	expectedSub := fromBigInt(bigFromHex("67b381821c52a5624f3304a8149be8461e3bc07adcb872c38aa65051ba53ba97"))
-	expectedNeg := fromBigInt(bigFromHex("7f1d8aad70909be90358f1d02240062433cc3a0248ded72febb879ec33ce6f22"))
-	expectedMul := fromBigInt(bigFromHex("3d08bbad376584e4f74bd31f78f716372b96ba8c3f939c12b8d54e79b6489e76"))
-	expectedMul2 := fromBigInt(bigFromHex("1df94a9e05a559ff38e0ab50cece734dc058d33738ceacaa15986a67cbff1ef6"))
-
-	t.Parallel()
-	t.Run("add", func(t *testing.T) {
-		ret := &gfP{}
-		gfpAddAsm(ret, x, y)
-		if *expectedAdd != *ret {
-			t.Errorf("add not same")
-		}
-		x1 := &gfP{}
-		x1.Set(x)
-		gfpAddAsm(x1, x1, y)
-		if *expectedAdd != *x1 {
-			t.Errorf("add not same when add self")
-		}
-	})
-
-	t.Run("double", func(t *testing.T) {
-		ret := &gfP{}
-		gfpDoubleAsm(ret, x)
-		if ret.Equal(expectedDouble) != 1 {
-			t.Errorf("double not same, got %v, expected %v", ret, expectedDouble)
-		}
-		ret.Set(x)
-		gfpDoubleAsm(ret, ret)
-		if ret.Equal(expectedDouble) != 1 {
-			t.Errorf("double not same, got %v, expected %v", ret, expectedDouble)
-		}
-	})
-
-	t.Run("triple", func(t *testing.T) {
-		expected := &gfP{}
-		gfpAddAsm(expected, x, expectedDouble)
-		ret := &gfP{}
-		ret.Set(x)
-		gfpTripleAsm(ret, ret)
-		if ret.Equal(expected) != 1 {
-			t.Errorf("expected %v, got %v", expected, ret)
-		}
-	})
-
-	t.Run("sub", func(t *testing.T) {
-		ret := &gfP{}
-		gfpSubAsm(ret, y, x)
-		if *expectedSub != *ret {
-			t.Errorf("sub not same")
-		}
-		x1 := &gfP{}
-		x1.Set(x)
-		gfpSubAsm(x1, y, x1)
-		if *expectedSub != *x1 {
-			t.Errorf("sub not same when sub self")
-		}
-		gfpSubAsm(ret, x, x)
-		if *ret != *zero {
-			t.Errorf("expected zero")
-		}
-	})
-
-	t.Run("neg", func(t *testing.T) {
-		ret := &gfP{}
-		gfpNegAsm(ret, y)
-		if *expectedNeg != *ret {
-			t.Errorf("neg not same")
-		}
-		ret.Set(y)
-		gfpNegAsm(ret, ret)
-		if *expectedNeg != *ret {
-			t.Errorf("neg not same when neg self")
-		}
-	})
-
-	t.Run("mul", func(t *testing.T) {
-		ret := &gfP{}
-		gfpMulAsm(ret, x, y)
-		if *expectedMul != *ret {
-			t.Errorf("mul not same")
-		}
-		ret.Set(x)
-		gfpMulAsm(ret, ret, y)
-		if *expectedMul != *ret {
-			t.Errorf("mul not same when mul self")
-		}
-	})
-
-	t.Run("square", func(t *testing.T) {
-		ret, ret1, ret2 := &gfP{}, &gfP{}, &gfP{}
-		gfpMulAsm(ret, x, y)
-		gfpMulAsm(ret1, ret, ret)
-		if *ret1 != *expectedMul2 {
-			t.Errorf("mul not same")
-		}
-		gfpMulAsm(ret1, ret1, ret1)
-		gfpSqrAsm(ret2, ret, 2)
-		if *ret1 != *ret2 {
-			t.Errorf("mul/sqr not same")
-		}
-		ret2.Set(ret)
-		gfpSqrAsm(ret2, ret2, 2)
-		if *ret1 != *ret2 {
-			t.Errorf("mul/sqr not same when square self")
-		}
-	})
-}