From b7873f91096f6b6759afbdfe032d13aab9e9075c Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Fri, 22 Jul 2016 15:16:21 +0900
Subject: [PATCH] add test of mulPre768(not run)

---
 Makefile         |  2 +-
 sample/large.cpp | 45 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 46 insertions(+), 1 deletion(-)

diff --git a/Makefile b/Makefile
index 6d7d3c8..30ff027 100644
--- a/Makefile
+++ b/Makefile
@@ -34,7 +34,7 @@ LLVM_SRC=src/base$(BIT).ll
 # CPU is used for llvm
 # see $(LLVM_LLC) --version
 LLVM_FLAGS=-march=$(CPU) -relocation-model=pic #-misched=ilpmax
-#LLVM_FLAGS+=-pre-RA-sched=list-ilp -max-sched-reorder=128
+LLVM_FLAGS+=-pre-RA-sched=list-ilp -max-sched-reorder=128
 
 HAS_BMI2=$(shell cat "/proc/cpuinfo" | grep bmi2 >/dev/null && echo "1")
 ifeq ($(HAS_BMI2),1)
diff --git a/sample/large.cpp b/sample/large.cpp
index 5397697..5a6e37a 100644
--- a/sample/large.cpp
+++ b/sample/large.cpp
@@ -8,7 +8,33 @@
 typedef mcl::FpT<> Fp;
 
 typedef mcl::fp::Unit Unit;
+using namespace mcl::fp;
 
+#include "../src/low_gmp.hpp"
+const size_t N = 12;
+
+void mulPre768(Unit *pz, const Unit *px, const Unit *py)
+{
+	/*
+		W = 1 << H
+		(aW + b)(cW + d) = acW^2 + (ad + bc)W + bd
+		ad + bc = (a + b)(c + d) - ac - bd
+	*/
+	const size_t H = N / 2;
+	low_mul<H>(pz, px, py); // bd
+	low_mul<H>(pz + N, px + H, py + H); // ac
+	Unit a_b[H + 1];
+	Unit c_d[H + 1];
+	a_b[H] = low_add<H>(a_b, px, px + H); // a + b
+	c_d[H] = low_add<H>(c_d, py, py + H); // c + d
+	Unit work[N + H] = {};
+	low_mul<H>(work, a_b, c_d);
+	if (c_d[H]) low_add<H + 1>(work + H, work + H, c_d);
+	if (a_b[H]) low_add<H + 1>(work + H, work + H, a_b);
+	work[N] -= low_sub<H>(work, work, pz);
+	work[N] -= low_sub<H>(work, work, pz + N);
+	low_add<H + N>(pz + H, pz + H, work);
+}
 void test(const std::string& pStr, mcl::fp::Mode mode)
 {
 	printf("test %s\n", mcl::fp::ModeToStr(mode));
@@ -31,6 +57,24 @@ void test(const std::string& pStr, mcl::fp::Mode mode)
 	CYBOZU_BENCH("mul", Fp::mul, x, x, x);
 }
 
+void testMul()
+{
+	mcl::fp::Unit ux[N], uy[N], a[N * 2], b[N * 2];
+	for (size_t i = 0; i < N; i++) {
+		ux[i] = -i * i + 5;
+		uy[i] = -i * i + 9;
+	}
+	low_mul<12>(a, ux, uy);
+	mulPre768(b, ux, uy);
+	for (size_t i = 0; i < N * 2; i++) {
+		if (a[i] != b[i]) {
+			printf("ERR %016llx %016llx\n", (long long)a[i], (long long)b[i]);
+		}
+	}
+	puts("end testMul");
+	CYBOZU_BENCH("mulPre768", mulPre768, ux, ux, uy);
+}
+
 void testAll(const std::string& pStr)
 {
 	test(pStr, mcl::fp::FP_GMP);
@@ -49,6 +93,7 @@ int main()
 	};
 	testAll(pTbl[0]);
 	testAll(pTbl[1]);
+//	testMul();
 } catch (std::exception& e) {
 	printf("err %s\n", e.what());
 	puts("make clean");