From c09de3ee3a303bfab596def8e0b5c8b845e5a97f Mon Sep 17 00:00:00 2001
From: Rob Austein <sra@hactrn.net>
Date: Tue, 7 Mar 2017 19:55:09 -0500
Subject: Promote to a repository in the core tree.

Change name of reset signal from rst_n to reset_n for consistancy with
other Cryptech cores.

Code common between this core and the ecdsa256 core split out into a
separate library repository.

Minor cleanup (Windows-isms, indentation).
---
 bench/tb_curve_adder_384.v        | 750 +++++++++++++++++++-------------------
 bench/tb_curve_doubler_384.v      | 724 ++++++++++++++++++------------------
 bench/tb_curve_multiplier_384.v   | 460 +++++++++++------------
 bench/tb_modular_multiplier_384.v | 648 ++++++++++++++++----------------
 4 files changed, 1291 insertions(+), 1291 deletions(-)

(limited to 'bench')

diff --git a/bench/tb_curve_adder_384.v b/bench/tb_curve_adder_384.v
index b2da029..4fee4e1 100644
--- a/bench/tb_curve_adder_384.v
+++ b/bench/tb_curve_adder_384.v
@@ -43,383 +43,383 @@
 module tb_curve_adder_384;
 
 
-		//
-		// Test Vectors
-		//
-	localparam	[383:0]	PX_1	= 384'h7d51764067faaac686ee307807af5544a93e13c540cac538c7853a590102fa5fe6cdcc7791e44b76ef21c57e21df37e5;
-	localparam	[383:0]	PY_1	= 384'h046ed01f219649209a4cfd43572d7bdd1f10f74d5be895a6c8da8e8edfc2601aaf7651e497b8688cf02ba5e1c7e77773;
-	localparam	[383:0]	PZ_1	= 384'h27c452f26af1ece2924446b574969fd102556bd28712ff4b3eb1044c5ec23e59f958793e10a510f0aef98c3b724df1d5;
-
-	localparam	[383:0]	RX_1	= 384'h238faca9e58d1fe1b59a24ba94ec6d9836fc360cb84e103715f1761554ff6a09b6d605a34f63ae0f995f59c162e7ff24;
-	localparam	[383:0]	RY_1	= 384'h63b7c7672a310f779a2047315523788d69b823d4e97e26c3b45201b0345a95de977024e97e8648215a637727d0f17747;
-	localparam	[383:0]	RZ_1	= 384'h394b5d916df6f9120c7771c750df3bc910998eb0a08daac1ca6e15ef70780fd48bf794d06f9cbe9568a2dbf4362dda86;
-
-
-	localparam	[383:0]	PX_2	= 384'h8418363a2fe99e888abba4df9e4a0e55452b9e968454ffadc96b4fb8072174109755c564a7be3c2f860652315d635f56;
-	localparam	[383:0]	PY_2	= 384'h2bef570ce39347040330df4ac581fcc7d9dd9deb286bd80f05257d90d6560f1b5381009a3a6f0acc1ea30a5e7cc7d8b8;
-	localparam	[383:0]	PZ_2	= 384'h86a8b840bd52d4dc58e9e4323fd4b40ea9b262a8cb45f7e95f5407b7e5eacc16fe4ce70125f20b76b37900af12cbf909;
-
-	localparam	[383:0]	RX_2	= 384'h7de41f9c2d48cd65d2cd1288c4fcdcd5bfe37575b23b8784e6091917a1c92c264a1105ce3c4ab88ca53947f35610f671;
-	localparam	[383:0]	RY_2	= 384'h419decfcef28d06a824595bbcb86ff56aebe48a33ceb80f0a256b90aee8214d0d879454842457d49e00a2330cc0ccb57;
-	localparam	[383:0]	RZ_2	= 384'h63843f81a3427e4dea2e3e150c16d26a4aa2af4d8037a1012e0490babdfa55606e5cea5e183cc147f2c1b377748ae58b;
-
-
-	localparam	[383:0]	PX_3	= 384'hd95cd4004f9417b3bea71ac087da4849841c4fbb41b500736773f41f241414024d0025e1a4337e5b24b1fe2ba2e51a83;
-	localparam	[383:0]	PY_3	= 384'h53650c78f3347a07248d6f6452dd80f5e373c7eddc810c6123c912b55da1a297e147e834547d1af4938fc139e71958f7;
-	localparam	[383:0]	PZ_3	= 384'hfbcfa387106d61d1e1e5f660206189f13e0861b3eb2e880b200cae1878b8b47bd5b727b2825a5f395025bf42a571093b;
-
-	localparam	[383:0]	RX_3	= 384'hf94e9fb67d6bac22276bb4fb4f965419f1907078236dd8520ff416e8525933055795c496d344a2e1cce480557dde88f4;
-	localparam	[383:0]	RY_3	= 384'h3feb5bc25c380ead1a149ab2f5e3ac4cfeaa5d91bd66dda0f29d6e02520ea8583bb3ab744fb826c18981f7bac03a0886;
-	localparam	[383:0]	RZ_3	= 384'hbc364ce11866cdeb1abdbc34b633e54fbb000090134600cb6dbff71f69e3d24843e40af30c07444530b53acb89eef49b;
-
-
-	localparam	[383:0]	PX_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
-	localparam	[383:0]	PY_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
-	localparam	[383:0]	PZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
-
-	localparam	[383:0]	RX_4	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7; // G.x
-	localparam	[383:0]	RY_4	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f; // G.y
-	localparam	[383:0]	RZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
-
-
-	localparam	[383:0]	PX_5	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7; // G.x
-	localparam	[383:0]	PY_5	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f; // G.y
-	localparam	[383:0]	PZ_5	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
-
-	localparam	[383:0]	RX_5	= 384'haaf06bba82e9f590e29c71c219bea51723c5893ae8b0c8cf4c117c3efb57ab8d55fa1b428155ad278b5743911b13ea8a; // H.x
-	localparam	[383:0]	RY_5	= 384'hc9e821b569d9d390a26167406d6d23d6070be242d765eb831625ceec4a0f473ef59f4e30e2817e6285bce2846f15f19d; // H.y
-	localparam	[383:0]	RZ_5	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
-
-	
-	localparam	[383:0]	PX_6	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7; // G.x
-	localparam	[383:0]	PY_6	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f; // G.y
-	localparam	[383:0]	PZ_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
-
-	localparam	[383:0]	RX_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;	//
-	localparam	[383:0]	RY_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;	// O
-	localparam	[383:0]	RZ_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;	//
-
-
-	localparam	[383:0]	Q		= 384'hfffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffeffffffff0000000000000000ffffffff;
-		
-		
-		//
-		// Core Parameters
-		//
-	localparam	WORD_COUNTER_WIDTH	=  4;
-	localparam	OPERAND_NUM_WORDS		= 12;
-	
-
-		//
-		// Clock (100 MHz)
-		//
-	reg clk = 1'b0;
-	always #5 clk = ~clk;
-
-	
-		//
-		// Inputs, Outputs
-		//
-	reg	rst_n;
-	reg	ena;
-	wire	rdy;
-
-	
-		//
-		// Buffers (PX, PY, PZ, RX, RY, RZ, Q)
-		//
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_px_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_py_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_pz_addr;
-	
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_rx_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_ry_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_rz_addr;
-	
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_q_addr;
-	
-	wire										core_rx_wren;
-	wire										core_ry_wren;
-	wire										core_rz_wren;
-	
-	wire	[                32-1:0]	core_px_data;
-	wire	[                32-1:0]	core_py_data;
-	wire	[                32-1:0]	core_pz_data;
-	
-	wire	[                32-1:0]	core_rx_data_wr;
-	wire	[                32-1:0]	core_ry_data_wr;
-	wire	[                32-1:0]	core_rz_data_wr;
-	
-	wire	[                32-1:0]	core_rx_data_rd;
-	wire	[                32-1:0]	core_ry_data_rd;
-	wire	[                32-1:0]	core_rz_data_rd;	
-	
-	wire	[                32-1:0]	core_q_data;
-	
-	reg	[WORD_COUNTER_WIDTH-1:0]	tb_xyzq_addr;
-	reg										tb_xyzq_wren;
-	
-	reg	[                  31:0]	tb_px_data;
-	reg	[                  31:0]	tb_py_data;
-	reg	[                  31:0]	tb_pz_data;
-	wire	[                  31:0]	tb_rx_data;
-	wire	[                  31:0]	tb_ry_data;
-	wire	[                  31:0]	tb_rz_data;
-	reg	[                  31:0]	tb_q_data;
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_px
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_px_data), .a_out(),
-		.b_addr(core_px_addr), .b_out(core_px_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_py
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_py_data), .a_out(),
-		.b_addr(core_py_addr), .b_out(core_py_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_pz
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_pz_data), .a_out(),
-		.b_addr(core_pz_addr), .b_out(core_pz_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_q
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_q_data), .a_out(),
-		.b_addr(core_q_addr), .b_out(core_q_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_rx
-	(	.clk(clk),
-		.a_addr(core_rx_addr), .a_wr(core_rx_wren), .a_in(core_rx_data_wr), .a_out(core_rx_data_rd),
-		.b_addr(tb_xyzq_addr), .b_out(tb_rx_data)
-	);	
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_ry
-	(	.clk(clk),
-		.a_addr(core_ry_addr), .a_wr(core_ry_wren), .a_in(core_ry_data_wr), .a_out(core_ry_data_rd),
-		.b_addr(tb_xyzq_addr), .b_out(tb_ry_data)
-	);	
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_rz
-	(	.clk(clk),
-		.a_addr(core_rz_addr), .a_wr(core_rz_wren), .a_in(core_rz_data_wr), .a_out(core_rz_data_rd),
-		.b_addr(tb_xyzq_addr), .b_out(tb_rz_data)
-	);
-	
-	
-		//
-		// Opcode
-		//
-	wire	[ 5: 0]	add_uop_addr;
-	wire	[19: 0]	add_uop;
-	
-	uop_add_rom add_rom
-	(
-		.clk		(clk),
-		.addr		(add_uop_addr),
-		.data		(add_uop)
-    );
-	
-		//
-		// UUT
-		//
-	curve_dbl_add_384 uut
-	(
-		.clk		(clk),
-		.rst_n	(rst_n),
-		
-		.ena		(ena),
-		.rdy		(rdy),
-		
-		.uop_addr	(add_uop_addr),
-		.uop			(add_uop),
-
-		.px_addr	(core_px_addr),
-		.py_addr	(core_py_addr),
-		.pz_addr	(core_pz_addr),
-		.rx_addr	(core_rx_addr),
-		.ry_addr	(core_ry_addr),
-		.rz_addr	(core_rz_addr),
-		.q_addr	(core_q_addr),
-		
-		.rx_wren	(core_rx_wren),
-		.ry_wren	(core_ry_wren),
-		.rz_wren	(core_rz_wren),
-		
-		.px_din	(core_px_data),
-		.py_din	(core_py_data),
-		.pz_din	(core_pz_data),
-		.rx_din	(core_rx_data_rd),
-		.ry_din	(core_ry_data_rd),
-		.rz_din	(core_rz_data_rd),		
-		.rx_dout	(core_rx_data_wr),
-		.ry_dout	(core_ry_data_wr),
-		.rz_dout	(core_rz_data_wr),
-		.q_din	(core_q_data)
+   //
+   // Test Vectors
+   //
+   localparam	[383:0]	PX_1	= 384'h7d51764067faaac686ee307807af5544a93e13c540cac538c7853a590102fa5fe6cdcc7791e44b76ef21c57e21df37e5;
+   localparam	[383:0]	PY_1	= 384'h046ed01f219649209a4cfd43572d7bdd1f10f74d5be895a6c8da8e8edfc2601aaf7651e497b8688cf02ba5e1c7e77773;
+   localparam	[383:0]	PZ_1	= 384'h27c452f26af1ece2924446b574969fd102556bd28712ff4b3eb1044c5ec23e59f958793e10a510f0aef98c3b724df1d5;
+
+   localparam	[383:0]	RX_1	= 384'h238faca9e58d1fe1b59a24ba94ec6d9836fc360cb84e103715f1761554ff6a09b6d605a34f63ae0f995f59c162e7ff24;
+   localparam	[383:0]	RY_1	= 384'h63b7c7672a310f779a2047315523788d69b823d4e97e26c3b45201b0345a95de977024e97e8648215a637727d0f17747;
+   localparam	[383:0]	RZ_1	= 384'h394b5d916df6f9120c7771c750df3bc910998eb0a08daac1ca6e15ef70780fd48bf794d06f9cbe9568a2dbf4362dda86;
+
+
+   localparam	[383:0]	PX_2	= 384'h8418363a2fe99e888abba4df9e4a0e55452b9e968454ffadc96b4fb8072174109755c564a7be3c2f860652315d635f56;
+   localparam	[383:0]	PY_2	= 384'h2bef570ce39347040330df4ac581fcc7d9dd9deb286bd80f05257d90d6560f1b5381009a3a6f0acc1ea30a5e7cc7d8b8;
+   localparam	[383:0]	PZ_2	= 384'h86a8b840bd52d4dc58e9e4323fd4b40ea9b262a8cb45f7e95f5407b7e5eacc16fe4ce70125f20b76b37900af12cbf909;
+
+   localparam	[383:0]	RX_2	= 384'h7de41f9c2d48cd65d2cd1288c4fcdcd5bfe37575b23b8784e6091917a1c92c264a1105ce3c4ab88ca53947f35610f671;
+   localparam	[383:0]	RY_2	= 384'h419decfcef28d06a824595bbcb86ff56aebe48a33ceb80f0a256b90aee8214d0d879454842457d49e00a2330cc0ccb57;
+   localparam	[383:0]	RZ_2	= 384'h63843f81a3427e4dea2e3e150c16d26a4aa2af4d8037a1012e0490babdfa55606e5cea5e183cc147f2c1b377748ae58b;
+
+
+   localparam	[383:0]	PX_3	= 384'hd95cd4004f9417b3bea71ac087da4849841c4fbb41b500736773f41f241414024d0025e1a4337e5b24b1fe2ba2e51a83;
+   localparam	[383:0]	PY_3	= 384'h53650c78f3347a07248d6f6452dd80f5e373c7eddc810c6123c912b55da1a297e147e834547d1af4938fc139e71958f7;
+   localparam	[383:0]	PZ_3	= 384'hfbcfa387106d61d1e1e5f660206189f13e0861b3eb2e880b200cae1878b8b47bd5b727b2825a5f395025bf42a571093b;
+
+   localparam	[383:0]	RX_3	= 384'hf94e9fb67d6bac22276bb4fb4f965419f1907078236dd8520ff416e8525933055795c496d344a2e1cce480557dde88f4;
+   localparam	[383:0]	RY_3	= 384'h3feb5bc25c380ead1a149ab2f5e3ac4cfeaa5d91bd66dda0f29d6e02520ea8583bb3ab744fb826c18981f7bac03a0886;
+   localparam	[383:0]	RZ_3	= 384'hbc364ce11866cdeb1abdbc34b633e54fbb000090134600cb6dbff71f69e3d24843e40af30c07444530b53acb89eef49b;
+
+
+   localparam	[383:0]	PX_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
+   localparam	[383:0]	PY_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
+   localparam	[383:0]	PZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
+
+   localparam	[383:0]	RX_4	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7; // G.x
+   localparam	[383:0]	RY_4	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f; // G.y
+   localparam	[383:0]	RZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
+
+
+   localparam	[383:0]	PX_5	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7; // G.x
+   localparam	[383:0]	PY_5	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f; // G.y
+   localparam	[383:0]	PZ_5	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
+
+   localparam	[383:0]	RX_5	= 384'haaf06bba82e9f590e29c71c219bea51723c5893ae8b0c8cf4c117c3efb57ab8d55fa1b428155ad278b5743911b13ea8a; // H.x
+   localparam	[383:0]	RY_5	= 384'hc9e821b569d9d390a26167406d6d23d6070be242d765eb831625ceec4a0f473ef59f4e30e2817e6285bce2846f15f19d; // H.y
+   localparam	[383:0]	RZ_5	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
+
+
+   localparam	[383:0]	PX_6	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7; // G.x
+   localparam	[383:0]	PY_6	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f; // G.y
+   localparam	[383:0]	PZ_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
+
+   localparam	[383:0]	RX_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;	//
+   localparam	[383:0]	RY_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;	// O
+   localparam	[383:0]	RZ_6	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;	//
+
+
+   localparam	[383:0]	Q	= 384'hfffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffeffffffff0000000000000000ffffffff;
+
+
+   //
+   // Core Parameters
+   //
+   localparam	WORD_COUNTER_WIDTH	=  4;
+   localparam	OPERAND_NUM_WORDS	= 12;
+
+
+   //
+   // Clock (100 MHz)
+   //
+   reg clk = 1'b0;
+   always #5 clk = ~clk;
+
+
+   //
+   // Inputs, Outputs
+   //
+   reg rst_n;
+   reg ena;
+   wire rdy;
+
+
+   //
+   // Buffers (PX, PY, PZ, RX, RY, RZ, Q)
+   //
+   wire [WORD_COUNTER_WIDTH-1:0] core_px_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_py_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_pz_addr;
+
+   wire [WORD_COUNTER_WIDTH-1:0] core_rx_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_ry_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_rz_addr;
+
+   wire [WORD_COUNTER_WIDTH-1:0] core_q_addr;
+
+   wire 			 core_rx_wren;
+   wire 			 core_ry_wren;
+   wire 			 core_rz_wren;
+
+   wire [                32-1:0] core_px_data;
+   wire [                32-1:0] core_py_data;
+   wire [                32-1:0] core_pz_data;
+
+   wire [                32-1:0] core_rx_data_wr;
+   wire [                32-1:0] core_ry_data_wr;
+   wire [                32-1:0] core_rz_data_wr;
+
+   wire [                32-1:0] core_rx_data_rd;
+   wire [                32-1:0] core_ry_data_rd;
+   wire [                32-1:0] core_rz_data_rd;
+
+   wire [                32-1:0] core_q_data;
+
+   reg [WORD_COUNTER_WIDTH-1:0]  tb_xyzq_addr;
+   reg 				 tb_xyzq_wren;
+
+   reg [                  31:0]  tb_px_data;
+   reg [                  31:0]  tb_py_data;
+   reg [                  31:0]  tb_pz_data;
+   wire [                  31:0] tb_rx_data;
+   wire [                  31:0] tb_ry_data;
+   wire [                  31:0] tb_rz_data;
+   reg [                  31:0]  tb_q_data;
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_px
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_px_data), .a_out(),
+	.b_addr(core_px_addr), .b_out(core_px_data)
 	);
-		
-		
-		//
-		// Testbench Routine
-		//
-	reg ok = 1;
-	initial begin
-		
-			/* initialize control inputs */
-		rst_n		= 0;
-		ena		= 0;
-		
-			/* wait for some time */
-		#200;
-		
-			/* de-assert reset */
-		rst_n		= 1;
-		
-			/* wait for some time */
-		#100;		
-		
-			/* run tests */
-		test_curve_adder(PX_1,     PY_1, PZ_1, RX_1, RY_1, RZ_1);
-		test_curve_adder(PX_2,     PY_2, PZ_2, RX_2, RY_2, RZ_2);
-		test_curve_adder(PX_3,     PY_3, PZ_3, RX_3, RY_3, RZ_3);
-		test_curve_adder(PX_4,     PY_4, PZ_4, RX_4, RY_4, RZ_4);
-		test_curve_adder(PX_5,     PY_5, PZ_5, RX_5, RY_5, RZ_5);
-		test_curve_adder(PX_6, Q - PY_6, PZ_6, RX_6, RY_6, RZ_6);
-		
-			/* print result */
-		if (ok)	$display("tb_curve_adder_384: SUCCESS");
-		else		$display("tb_curve_adder_384: FAILURE");
-		//
-		$finish;
-		//
-	end
-	
-	
-		//
-		// Test Task
-		//	
-	reg		t_ok;
-	
-	integer	w;
-
-	task test_curve_adder;
-	
-		input	[383:0]	px;
-		input	[383:0]	py;
-		input	[383:0]	pz;
-		
-		input	[383:0]	rx;
-		input	[383:0]	ry;
-		input	[383:0]	rz;		
-		
-		reg	[383:0]	px_shreg;
-		reg	[383:0]	py_shreg;
-		reg	[383:0]	pz_shreg;
-		
-		reg	[383:0]	rx_shreg;
-		reg	[383:0]	ry_shreg;
-		reg	[383:0]	rz_shreg;
-		
-		reg	[383:0]	q_shreg;
-		
-		begin
-		
-				/* start filling memories */
-			tb_xyzq_wren = 1;
-			
-				/* initialize shift registers */
-			px_shreg = px;
-			py_shreg = py;
-			pz_shreg = pz;
-			q_shreg  = Q;
-			
-				/* write all the words */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set addresses */
-				tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* set data words */
-				tb_px_data	= px_shreg[31:0];
-				tb_py_data	= py_shreg[31:0];
-				tb_pz_data	= pz_shreg[31:0];
-				tb_q_data	= q_shreg[31:0];
-				
-					/* shift inputs */
-				px_shreg = {{32{1'bX}}, px_shreg[383:32]};
-				py_shreg = {{32{1'bX}}, py_shreg[383:32]};
-				pz_shreg = {{32{1'bX}}, pz_shreg[383:32]};
-				q_shreg  = {{32{1'bX}}, q_shreg[383:32]};
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-			end
-			
-				/* wipe addresses */
-			tb_xyzq_addr = {WORD_COUNTER_WIDTH{1'bX}};
-			
-				/* wipe data words */
-			tb_px_data = {32{1'bX}};
-			tb_py_data = {32{1'bX}};
-			tb_pz_data = {32{1'bX}};
-			tb_q_data  = {32{1'bX}};
-			
-				/* stop filling memories */
-			tb_xyzq_wren = 0;
-			
-				/* start operation */
-			ena = 1;
-			
-				/* clear flag */
-			#10 ena = 0;
-			
-				/* wait for operation to complete */
-			while (!rdy) #10;
-			
-				/* read result */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set address */
-				tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-					/* store data word */
-				rx_shreg = {tb_rx_data, rx_shreg[383:32]};
-				ry_shreg = {tb_ry_data, ry_shreg[383:32]};
-				rz_shreg = {tb_rz_data, rz_shreg[383:32]};
-
-			end
-			
-				/* compare */
-			t_ok =	(rx_shreg == rx) &&
-						(ry_shreg == ry) &&
-						(rz_shreg == rz);
-
-				/* display results */
-			$display("test_curve_adder(): %s", t_ok ? "OK" : "ERROR");
-			
-				/* update global flag */
-			ok = ok && t_ok;
-		
-		end
-		
-	endtask
-	
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_py
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_py_data), .a_out(),
+	.b_addr(core_py_addr), .b_out(core_py_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_pz
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_pz_data), .a_out(),
+	.b_addr(core_pz_addr), .b_out(core_pz_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_q
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_q_data), .a_out(),
+	.b_addr(core_q_addr), .b_out(core_q_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_rx
+     (	.clk(clk),
+	.a_addr(core_rx_addr), .a_wr(core_rx_wren), .a_in(core_rx_data_wr), .a_out(core_rx_data_rd),
+	.b_addr(tb_xyzq_addr), .b_out(tb_rx_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_ry
+     (	.clk(clk),
+	.a_addr(core_ry_addr), .a_wr(core_ry_wren), .a_in(core_ry_data_wr), .a_out(core_ry_data_rd),
+	.b_addr(tb_xyzq_addr), .b_out(tb_ry_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_rz
+     (	.clk(clk),
+	.a_addr(core_rz_addr), .a_wr(core_rz_wren), .a_in(core_rz_data_wr), .a_out(core_rz_data_rd),
+	.b_addr(tb_xyzq_addr), .b_out(tb_rz_data)
+	);
+
+
+   //
+   // Opcode
+   //
+   wire [ 5: 0] 		 add_uop_addr;
+   wire [19: 0] 		 add_uop;
+
+   uop_add_rom add_rom
+     (
+      .clk		(clk),
+      .addr		(add_uop_addr),
+      .data		(add_uop)
+      );
+
+   //
+   // UUT
+   //
+   curve_dbl_add_384 uut
+     (
+      .clk		(clk),
+      .rst_n	(rst_n),
+
+      .ena		(ena),
+      .rdy		(rdy),
+
+      .uop_addr	(add_uop_addr),
+      .uop			(add_uop),
+
+      .px_addr	(core_px_addr),
+      .py_addr	(core_py_addr),
+      .pz_addr	(core_pz_addr),
+      .rx_addr	(core_rx_addr),
+      .ry_addr	(core_ry_addr),
+      .rz_addr	(core_rz_addr),
+      .q_addr	(core_q_addr),
+
+      .rx_wren	(core_rx_wren),
+      .ry_wren	(core_ry_wren),
+      .rz_wren	(core_rz_wren),
+
+      .px_din	(core_px_data),
+      .py_din	(core_py_data),
+      .pz_din	(core_pz_data),
+      .rx_din	(core_rx_data_rd),
+      .ry_din	(core_ry_data_rd),
+      .rz_din	(core_rz_data_rd),
+      .rx_dout	(core_rx_data_wr),
+      .ry_dout	(core_ry_data_wr),
+      .rz_dout	(core_rz_data_wr),
+      .q_din	(core_q_data)
+      );
+
+
+   //
+   // Testbench Routine
+   //
+   reg 				 ok = 1;
+   initial begin
+
+      /* initialize control inputs */
+      rst_n		= 0;
+      ena		= 0;
+
+      /* wait for some time */
+      #200;
+
+      /* de-assert reset */
+      rst_n		= 1;
+
+      /* wait for some time */
+      #100;
+
+      /* run tests */
+      test_curve_adder(PX_1,     PY_1, PZ_1, RX_1, RY_1, RZ_1);
+      test_curve_adder(PX_2,     PY_2, PZ_2, RX_2, RY_2, RZ_2);
+      test_curve_adder(PX_3,     PY_3, PZ_3, RX_3, RY_3, RZ_3);
+      test_curve_adder(PX_4,     PY_4, PZ_4, RX_4, RY_4, RZ_4);
+      test_curve_adder(PX_5,     PY_5, PZ_5, RX_5, RY_5, RZ_5);
+      test_curve_adder(PX_6, Q - PY_6, PZ_6, RX_6, RY_6, RZ_6);
+
+      /* print result */
+      if (ok)	$display("tb_curve_adder_384: SUCCESS");
+      else	$display("tb_curve_adder_384: FAILURE");
+      //
+      $finish;
+      //
+   end
+
+
+   //
+   // Test Task
+   //
+   reg		t_ok;
+
+   integer	w;
+
+   task test_curve_adder;
+
+      input	[383:0]	px;
+      input [383:0] 	py;
+      input [383:0] 	pz;
+
+      input [383:0] 	rx;
+      input [383:0] 	ry;
+      input [383:0] 	rz;
+
+      reg [383:0] 	px_shreg;
+      reg [383:0] 	py_shreg;
+      reg [383:0] 	pz_shreg;
+
+      reg [383:0] 	rx_shreg;
+      reg [383:0] 	ry_shreg;
+      reg [383:0] 	rz_shreg;
+
+      reg [383:0] 	q_shreg;
+
+      begin
+
+	 /* start filling memories */
+	 tb_xyzq_wren = 1;
+
+	 /* initialize shift registers */
+	 px_shreg = px;
+	 py_shreg = py;
+	 pz_shreg = pz;
+	 q_shreg  = Q;
+
+	 /* write all the words */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set addresses */
+	    tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* set data words */
+	    tb_px_data	= px_shreg[31:0];
+	    tb_py_data	= py_shreg[31:0];
+	    tb_pz_data	= pz_shreg[31:0];
+	    tb_q_data	= q_shreg[31:0];
+
+	    /* shift inputs */
+	    px_shreg = {{32{1'bX}}, px_shreg[383:32]};
+	    py_shreg = {{32{1'bX}}, py_shreg[383:32]};
+	    pz_shreg = {{32{1'bX}}, pz_shreg[383:32]};
+	    q_shreg  = {{32{1'bX}}, q_shreg[383:32]};
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	 end
+
+	 /* wipe addresses */
+	 tb_xyzq_addr = {WORD_COUNTER_WIDTH{1'bX}};
+
+	 /* wipe data words */
+	 tb_px_data = {32{1'bX}};
+	 tb_py_data = {32{1'bX}};
+	 tb_pz_data = {32{1'bX}};
+	 tb_q_data  = {32{1'bX}};
+
+	 /* stop filling memories */
+	 tb_xyzq_wren = 0;
+
+	 /* start operation */
+	 ena = 1;
+
+	 /* clear flag */
+	 #10 ena = 0;
+
+	 /* wait for operation to complete */
+	 while (!rdy) #10;
+
+	 /* read result */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set address */
+	    tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	    /* store data word */
+	    rx_shreg = {tb_rx_data, rx_shreg[383:32]};
+	    ry_shreg = {tb_ry_data, ry_shreg[383:32]};
+	    rz_shreg = {tb_rz_data, rz_shreg[383:32]};
+
+	 end
+
+	 /* compare */
+	 t_ok =	(rx_shreg == rx) &&
+	       (ry_shreg == ry) &&
+	       (rz_shreg == rz);
+
+	 /* display results */
+	 $display("test_curve_adder(): %s", t_ok ? "OK" : "ERROR");
+
+	 /* update global flag */
+	 ok = ok && t_ok;
+
+      end
+
+   endtask
+
 endmodule
-
+
 
 //------------------------------------------------------------------------------
 // End-of-File
diff --git a/bench/tb_curve_doubler_384.v b/bench/tb_curve_doubler_384.v
index 30ef7ba..54a9914 100644
--- a/bench/tb_curve_doubler_384.v
+++ b/bench/tb_curve_doubler_384.v
@@ -43,369 +43,369 @@
 module tb_curve_doubler_384;
 
 
-		//
-		// Test Vectors
-		//
-	localparam	[383:0]	PX_1	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7;
-	localparam	[383:0]	PY_1	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f;
-	localparam	[383:0]	PZ_1	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
-
-	localparam	[383:0]	RX_1	= 384'he50dbe0981ef5f4f52eebd29e34e6d18d279318fe6b5d5616c54c93ea906c671b223d61ab7ba23cd59ab4f6ec5e40b94;
-	localparam	[383:0]	RY_1	= 384'h5b939b5a52ac7ebed90b8c1a809fb3a5c068b421b9e0a16208c2e53b1fe41e8373333e360ad2205e9dd63d29f1e0757a;
-	localparam	[383:0]	RZ_1	= 384'h6c2fbc952c4c58debb3d317f2525b853f1e83b7a513428f9d3b462276be1718014c1639c3afd033af4863af921d41cbe;
-
-
-	localparam	[383:0]	PX_2	= 384'heea14dc6c53e682b16c979fcbf2b39c0f5c43efa5b412f3e9bec8251a5ff9e243f46c6ba91a1604abdd5028b56e60334;
-	localparam	[383:0]	PY_2	= 384'h7ff6e6e9ed8da4aea5baa06d3a9583b6a3f206e935566659cd2025202a9e2a62e0c44e603f3a4304bfa974470f53f646;
-	localparam	[383:0]	PZ_2	= 384'hcb16156bf82550274e39900f2ed4359794160a166257ae2c71c6503c129b6ce92f31277387aa1c538e7702e3658a883d;
-
-	localparam	[383:0]	RX_2	= 384'h9455050d6d00285ae2cd25c95372b30d321fb31899f7b0e6f3b4dd557d3465edbc500cc403b076c7534a07a48d87824d;
-	localparam	[383:0]	RY_2	= 384'hd0cc7a7eb6fbe9cd962efa82f89b701fbd1f8a3579066feda96c9124ee3ef2764d923a2c0039e06865442dbf3ed460ae;
-	localparam	[383:0]	RZ_2	= 384'h473d297a1e5ed8b7afd51a27d5d933ab4555414ef75e1104c1fb56ba9b110f5d0d63a79e8e12edc4432c37663cc828f2;
-
-	
-	localparam	[383:0]	PX_3	= 384'h4af343df7a804f6b345562a471c4cea419f5e87086eaba95b7a7fa43aeb24a357d22047eea55c529fcdfb44ea80d3aab;
-	localparam	[383:0]	PY_3	= 384'hfa3e2ceb637c90fde564039da3e256456a8d0995c2f69846bf3e22f2f0760a2df00f2df2f79ce2484ed5b26124d733f1;
-	localparam	[383:0]	PZ_3	= 384'hdf144e3659137591d09ba17e7045f74c477e4f27e0f6c602f11306def2abeae08aee53fa40e3ca3b3ca52f9eaae47720;
-
-	localparam	[383:0]	RX_3	= 384'h9e2aad0ad19c585e03af779d1205a3a98bf7d367c0c829f741161983dc0d24d222d5d9c2ae39399ff40d4974df22ce4d;
-	localparam	[383:0]	RY_3	= 384'he0e40da86ae320396c813bb891f31f3d59be8c4957c602b56b5aa976e47d128127625b1ce7470fa33ed9e57eca5a268d;
-	localparam	[383:0]	RZ_3	= 384'h836e9d9117ba665abbee7063599a8631cd42f685a8bc29b05895c2928c19f640a065cfd2cc5b0d607ea52e388be513cf;
-
-	
-	localparam	[383:0]	PX_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
-	localparam	[383:0]	PY_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
-	localparam	[383:0]	PZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
-
-	localparam	[383:0]	RX_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
-	localparam	[383:0]	RY_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
-	localparam	[383:0]	RZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
-
-
-	localparam	[383:0]	Q		= 384'hfffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffeffffffff0000000000000000ffffffff;
-
-	
-	
-		//
-		// TODO: Test special cases!
-		//
-		
-		
-		//
-		// Core Parameters
-		//
-	localparam	WORD_COUNTER_WIDTH	=  4;
-	localparam	OPERAND_NUM_WORDS		= 12;
-	
-
-		//
-		// Clock (100 MHz)
-		//
-	reg clk = 1'b0;
-	always #5 clk = ~clk;
-
-	
-		//
-		// Inputs, Outputs
-		//
-	reg	rst_n;
-	reg	ena;
-	wire	rdy;
-
-	
-		//
-		// Buffers (PX, PY, PZ, RX, RY, RZ, Q)
-		//
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_px_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_py_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_pz_addr;
-	
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_rx_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_ry_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_rz_addr;
-	
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_q_addr;
-	
-	wire										core_rx_wren;
-	wire										core_ry_wren;
-	wire										core_rz_wren;
-	
-	wire	[                32-1:0]	core_px_data;
-	wire	[                32-1:0]	core_py_data;
-	wire	[                32-1:0]	core_pz_data;
-	
-	wire	[                32-1:0]	core_rx_data_wr;
-	wire	[                32-1:0]	core_ry_data_wr;
-	wire	[                32-1:0]	core_rz_data_wr;
-	
-	wire	[                32-1:0]	core_rx_data_rd;
-	wire	[                32-1:0]	core_ry_data_rd;
-	wire	[                32-1:0]	core_rz_data_rd;	
-	
-	wire	[                32-1:0]	core_q_data;
-	
-	reg	[WORD_COUNTER_WIDTH-1:0]	tb_xyzq_addr;
-	reg										tb_xyzq_wren;
-	
-	reg	[                  31:0]	tb_px_data;
-	reg	[                  31:0]	tb_py_data;
-	reg	[                  31:0]	tb_pz_data;
-	wire	[                  31:0]	tb_rx_data;
-	wire	[                  31:0]	tb_ry_data;
-	wire	[                  31:0]	tb_rz_data;
-	reg	[                  31:0]	tb_q_data;
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_px
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_px_data), .a_out(),
-		.b_addr(core_px_addr), .b_out(core_px_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_py
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_py_data), .a_out(),
-		.b_addr(core_py_addr), .b_out(core_py_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_pz
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_pz_data), .a_out(),
-		.b_addr(core_pz_addr), .b_out(core_pz_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_q
-	(	.clk(clk),
-		.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_q_data), .a_out(),
-		.b_addr(core_q_addr), .b_out(core_q_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_rx
-	(	.clk(clk),
-		.a_addr(core_rx_addr), .a_wr(core_rx_wren), .a_in(core_rx_data_wr), .a_out(core_rx_data_rd),
-		.b_addr(tb_xyzq_addr), .b_out(tb_rx_data)
-	);	
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_ry
-	(	.clk(clk),
-		.a_addr(core_ry_addr), .a_wr(core_ry_wren), .a_in(core_ry_data_wr), .a_out(core_ry_data_rd),
-		.b_addr(tb_xyzq_addr), .b_out(tb_ry_data)
-	);	
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_rz
-	(	.clk(clk),
-		.a_addr(core_rz_addr), .a_wr(core_rz_wren), .a_in(core_rz_data_wr), .a_out(core_rz_data_rd),
-		.b_addr(tb_xyzq_addr), .b_out(tb_rz_data)
-	);
-	
-	
-		//
-		// Opcode
-		//
-	wire	[ 5: 0]	dbl_uop_addr;
-	wire	[19: 0]	dbl_uop;
-	
-	uop_dbl_rom dbl_rom
-	(
-		.clk		(clk),
-		.addr		(dbl_uop_addr),
-		.data		(dbl_uop)
-    );
-	
-	
-		//
-		// UUT
-		//
-	curve_dbl_add_384 uut
-	(
-		.clk		(clk),
-		.rst_n	(rst_n),
-		
-		.ena		(ena),
-		.rdy		(rdy),
-		
-		.uop_addr	(dbl_uop_addr),
-		.uop			(dbl_uop),
-
-		.px_addr	(core_px_addr),
-		.py_addr	(core_py_addr),
-		.pz_addr	(core_pz_addr),
-		.rx_addr	(core_rx_addr),
-		.ry_addr	(core_ry_addr),
-		.rz_addr	(core_rz_addr),
-		.q_addr	(core_q_addr),
-		
-		.rx_wren	(core_rx_wren),
-		.ry_wren	(core_ry_wren),
-		.rz_wren	(core_rz_wren),
-		
-		.px_din	(core_px_data),
-		.py_din	(core_py_data),
-		.pz_din	(core_pz_data),
-		.rx_din	(core_rx_data_rd),
-		.ry_din	(core_ry_data_rd),
-		.rz_din	(core_rz_data_rd),		
-		.rx_dout	(core_rx_data_wr),
-		.ry_dout	(core_ry_data_wr),
-		.rz_dout	(core_rz_data_wr),
-		.q_din	(core_q_data)
+   //
+   // Test Vectors
+   //
+   localparam	[383:0]	PX_1	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7;
+   localparam	[383:0]	PY_1	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f;
+   localparam	[383:0]	PZ_1	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
+
+   localparam	[383:0]	RX_1	= 384'he50dbe0981ef5f4f52eebd29e34e6d18d279318fe6b5d5616c54c93ea906c671b223d61ab7ba23cd59ab4f6ec5e40b94;
+   localparam	[383:0]	RY_1	= 384'h5b939b5a52ac7ebed90b8c1a809fb3a5c068b421b9e0a16208c2e53b1fe41e8373333e360ad2205e9dd63d29f1e0757a;
+   localparam	[383:0]	RZ_1	= 384'h6c2fbc952c4c58debb3d317f2525b853f1e83b7a513428f9d3b462276be1718014c1639c3afd033af4863af921d41cbe;
+
+
+   localparam	[383:0]	PX_2	= 384'heea14dc6c53e682b16c979fcbf2b39c0f5c43efa5b412f3e9bec8251a5ff9e243f46c6ba91a1604abdd5028b56e60334;
+   localparam	[383:0]	PY_2	= 384'h7ff6e6e9ed8da4aea5baa06d3a9583b6a3f206e935566659cd2025202a9e2a62e0c44e603f3a4304bfa974470f53f646;
+   localparam	[383:0]	PZ_2	= 384'hcb16156bf82550274e39900f2ed4359794160a166257ae2c71c6503c129b6ce92f31277387aa1c538e7702e3658a883d;
+
+   localparam	[383:0]	RX_2	= 384'h9455050d6d00285ae2cd25c95372b30d321fb31899f7b0e6f3b4dd557d3465edbc500cc403b076c7534a07a48d87824d;
+   localparam	[383:0]	RY_2	= 384'hd0cc7a7eb6fbe9cd962efa82f89b701fbd1f8a3579066feda96c9124ee3ef2764d923a2c0039e06865442dbf3ed460ae;
+   localparam	[383:0]	RZ_2	= 384'h473d297a1e5ed8b7afd51a27d5d933ab4555414ef75e1104c1fb56ba9b110f5d0d63a79e8e12edc4432c37663cc828f2;
+
+
+   localparam	[383:0]	PX_3	= 384'h4af343df7a804f6b345562a471c4cea419f5e87086eaba95b7a7fa43aeb24a357d22047eea55c529fcdfb44ea80d3aab;
+   localparam	[383:0]	PY_3	= 384'hfa3e2ceb637c90fde564039da3e256456a8d0995c2f69846bf3e22f2f0760a2df00f2df2f79ce2484ed5b26124d733f1;
+   localparam	[383:0]	PZ_3	= 384'hdf144e3659137591d09ba17e7045f74c477e4f27e0f6c602f11306def2abeae08aee53fa40e3ca3b3ca52f9eaae47720;
+
+   localparam	[383:0]	RX_3	= 384'h9e2aad0ad19c585e03af779d1205a3a98bf7d367c0c829f741161983dc0d24d222d5d9c2ae39399ff40d4974df22ce4d;
+   localparam	[383:0]	RY_3	= 384'he0e40da86ae320396c813bb891f31f3d59be8c4957c602b56b5aa976e47d128127625b1ce7470fa33ed9e57eca5a268d;
+   localparam	[383:0]	RZ_3	= 384'h836e9d9117ba665abbee7063599a8631cd42f685a8bc29b05895c2928c19f640a065cfd2cc5b0d607ea52e388be513cf;
+
+
+   localparam	[383:0]	PX_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
+   localparam	[383:0]	PY_4	= 384'hxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx;
+   localparam	[383:0]	PZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
+
+   localparam	[383:0]	RX_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
+   localparam	[383:0]	RY_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001;
+   localparam	[383:0]	RZ_4	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
+
+
+   localparam	[383:0]	Q	= 384'hfffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffeffffffff0000000000000000ffffffff;
+
+
+
+   //
+   // TODO: Test special cases!
+   //
+
+
+   //
+   // Core Parameters
+   //
+   localparam	WORD_COUNTER_WIDTH	=  4;
+   localparam	OPERAND_NUM_WORDS	= 12;
+
+
+   //
+   // Clock (100 MHz)
+   //
+   reg clk = 1'b0;
+   always #5 clk = ~clk;
+
+
+   //
+   // Inputs, Outputs
+   //
+   reg rst_n;
+   reg ena;
+   wire rdy;
+
+
+   //
+   // Buffers (PX, PY, PZ, RX, RY, RZ, Q)
+   //
+   wire [WORD_COUNTER_WIDTH-1:0] core_px_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_py_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_pz_addr;
+
+   wire [WORD_COUNTER_WIDTH-1:0] core_rx_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_ry_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_rz_addr;
+
+   wire [WORD_COUNTER_WIDTH-1:0] core_q_addr;
+
+   wire 			 core_rx_wren;
+   wire 			 core_ry_wren;
+   wire 			 core_rz_wren;
+
+   wire [                32-1:0] core_px_data;
+   wire [                32-1:0] core_py_data;
+   wire [                32-1:0] core_pz_data;
+
+   wire [                32-1:0] core_rx_data_wr;
+   wire [                32-1:0] core_ry_data_wr;
+   wire [                32-1:0] core_rz_data_wr;
+
+   wire [                32-1:0] core_rx_data_rd;
+   wire [                32-1:0] core_ry_data_rd;
+   wire [                32-1:0] core_rz_data_rd;
+
+   wire [                32-1:0] core_q_data;
+
+   reg [WORD_COUNTER_WIDTH-1:0]  tb_xyzq_addr;
+   reg 				 tb_xyzq_wren;
+
+   reg [                  31:0]  tb_px_data;
+   reg [                  31:0]  tb_py_data;
+   reg [                  31:0]  tb_pz_data;
+   wire [                  31:0] tb_rx_data;
+   wire [                  31:0] tb_ry_data;
+   wire [                  31:0] tb_rz_data;
+   reg [                  31:0]  tb_q_data;
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_px
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_px_data), .a_out(),
+	.b_addr(core_px_addr), .b_out(core_px_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_py
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_py_data), .a_out(),
+	.b_addr(core_py_addr), .b_out(core_py_data)
 	);
-		
-		
-		//
-		// Testbench Routine
-		//
-	reg ok = 1;
-	initial begin
-		
-			/* initialize control inputs */
-		rst_n		= 0;
-		ena		= 0;
-		
-			/* wait for some time */
-		#200;
-		
-			/* de-assert reset */
-		rst_n		= 1;
-		
-			/* wait for some time */
-		#100;		
-		
-			/* run tests */
-		test_curve_doubler(PX_1, PY_1, PZ_1, RX_1, RY_1, RZ_1);
-		test_curve_doubler(PX_2, PY_2, PZ_2, RX_2, RY_2, RZ_2);
-		test_curve_doubler(PX_3, PY_3, PZ_3, RX_3, RY_3, RZ_3);
-		test_curve_doubler(PX_4, PY_4, PZ_4, RX_4, RY_4, RZ_4);
-		
-			/* print result */
-		if (ok)	$display("tb_curve_doubler_384: SUCCESS");
-		else		$display("tb_curve_doubler_384: FAILURE");
-		//
-//		$finish;
-		//
-	end
-	
-	
-		//
-		// Test Task
-		//	
-	reg		t_ok;
-	
-	integer	w;
-
-	task test_curve_doubler;
-	
-		input	[383:0]	px;
-		input	[383:0]	py;
-		input	[383:0]	pz;
-		
-		input	[383:0]	rx;
-		input	[383:0]	ry;
-		input	[383:0]	rz;		
-		
-		reg	[383:0]	px_shreg;
-		reg	[383:0]	py_shreg;
-		reg	[383:0]	pz_shreg;
-		
-		reg	[383:0]	rx_shreg;
-		reg	[383:0]	ry_shreg;
-		reg	[383:0]	rz_shreg;
-		
-		reg	[383:0]	q_shreg;
-		
-		begin
-		
-				/* start filling memories */
-			tb_xyzq_wren = 1;
-			
-				/* initialize shift registers */
-			px_shreg = px;
-			py_shreg = py;
-			pz_shreg = pz;
-			q_shreg  = Q;
-			
-				/* write all the words */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set addresses */
-				tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* set data words */
-				tb_px_data	= px_shreg[31:0];
-				tb_py_data	= py_shreg[31:0];
-				tb_pz_data	= pz_shreg[31:0];
-				tb_q_data	= q_shreg[31:0];
-				
-					/* shift inputs */
-				px_shreg = {{32{1'bX}}, px_shreg[383:32]};
-				py_shreg = {{32{1'bX}}, py_shreg[383:32]};
-				pz_shreg = {{32{1'bX}}, pz_shreg[383:32]};
-				q_shreg  = {{32{1'bX}}, q_shreg[383:32]};
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-			end
-			
-				/* wipe addresses */
-			tb_xyzq_addr = {WORD_COUNTER_WIDTH{1'bX}};
-			
-				/* wipe data words */
-			tb_px_data = {32{1'bX}};
-			tb_py_data = {32{1'bX}};
-			tb_pz_data = {32{1'bX}};
-			tb_q_data  = {32{1'bX}};
-			
-				/* stop filling memories */
-			tb_xyzq_wren = 0;
-			
-				/* start operation */
-			ena = 1;
-			
-				/* clear flag */
-			#10 ena = 0;
-			
-				/* wait for operation to complete */
-			while (!rdy) #10;
-			
-				/* read result */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set address */
-				tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-					/* store data word */
-				rx_shreg = {tb_rx_data, rx_shreg[383:32]};
-				ry_shreg = {tb_ry_data, ry_shreg[383:32]};
-				rz_shreg = {tb_rz_data, rz_shreg[383:32]};
-
-			end
-			
-				/* compare */
-			t_ok =	(rx_shreg == rx) &&
-						(ry_shreg == ry) &&
-						(rz_shreg == rz);
-
-				/* display results */
-			$display("test_curve_doubler(): %s", t_ok ? "OK" : "ERROR");
-			
-				/* update global flag */
-			ok = ok && t_ok;
-		
-		end
-		
-	endtask
-	
-      
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_pz
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_pz_data), .a_out(),
+	.b_addr(core_pz_addr), .b_out(core_pz_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_q
+     (	.clk(clk),
+	.a_addr(tb_xyzq_addr), .a_wr(tb_xyzq_wren), .a_in(tb_q_data), .a_out(),
+	.b_addr(core_q_addr), .b_out(core_q_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_rx
+     (	.clk(clk),
+	.a_addr(core_rx_addr), .a_wr(core_rx_wren), .a_in(core_rx_data_wr), .a_out(core_rx_data_rd),
+	.b_addr(tb_xyzq_addr), .b_out(tb_rx_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_ry
+     (	.clk(clk),
+	.a_addr(core_ry_addr), .a_wr(core_ry_wren), .a_in(core_ry_data_wr), .a_out(core_ry_data_rd),
+	.b_addr(tb_xyzq_addr), .b_out(tb_ry_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_rz
+     (	.clk(clk),
+	.a_addr(core_rz_addr), .a_wr(core_rz_wren), .a_in(core_rz_data_wr), .a_out(core_rz_data_rd),
+	.b_addr(tb_xyzq_addr), .b_out(tb_rz_data)
+	);
+
+
+   //
+   // Opcode
+   //
+   wire [ 5: 0]		 dbl_uop_addr;
+   wire [19: 0]		 dbl_uop;
+
+   uop_dbl_rom dbl_rom
+     (
+      .clk		(clk),
+      .addr		(dbl_uop_addr),
+      .data		(dbl_uop)
+      );
+
+
+   //
+   // UUT
+   //
+   curve_dbl_add_384 uut
+     (
+      .clk	(clk),
+      .rst_n	(rst_n),
+
+      .ena	(ena),
+      .rdy	(rdy),
+
+      .uop_addr	(dbl_uop_addr),
+      .uop			(dbl_uop),
+
+      .px_addr	(core_px_addr),
+      .py_addr	(core_py_addr),
+      .pz_addr	(core_pz_addr),
+      .rx_addr	(core_rx_addr),
+      .ry_addr	(core_ry_addr),
+      .rz_addr	(core_rz_addr),
+      .q_addr	(core_q_addr),
+
+      .rx_wren	(core_rx_wren),
+      .ry_wren	(core_ry_wren),
+      .rz_wren	(core_rz_wren),
+
+      .px_din	(core_px_data),
+      .py_din	(core_py_data),
+      .pz_din	(core_pz_data),
+      .rx_din	(core_rx_data_rd),
+      .ry_din	(core_ry_data_rd),
+      .rz_din	(core_rz_data_rd),
+      .rx_dout	(core_rx_data_wr),
+      .ry_dout	(core_ry_data_wr),
+      .rz_dout	(core_rz_data_wr),
+      .q_din	(core_q_data)
+      );
+
+
+   //
+   // Testbench Routine
+   //
+   reg 			 ok = 1;
+   initial begin
+
+      /* initialize control inputs */
+      rst_n		= 0;
+      ena		= 0;
+
+      /* wait for some time */
+      #200;
+
+      /* de-assert reset */
+      rst_n		= 1;
+
+      /* wait for some time */
+      #100;
+
+      /* run tests */
+      test_curve_doubler(PX_1, PY_1, PZ_1, RX_1, RY_1, RZ_1);
+      test_curve_doubler(PX_2, PY_2, PZ_2, RX_2, RY_2, RZ_2);
+      test_curve_doubler(PX_3, PY_3, PZ_3, RX_3, RY_3, RZ_3);
+      test_curve_doubler(PX_4, PY_4, PZ_4, RX_4, RY_4, RZ_4);
+
+      /* print result */
+      if (ok)	$display("tb_curve_doubler_384: SUCCESS");
+      else	$display("tb_curve_doubler_384: FAILURE");
+      //
+      //		$finish;
+      //
+   end
+
+
+   //
+   // Test Task
+   //
+   reg		t_ok;
+
+   integer	w;
+
+   task test_curve_doubler;
+
+      input	[383:0]	px;
+      input [383:0] 	py;
+      input [383:0] 	pz;
+
+      input [383:0] 	rx;
+      input [383:0] 	ry;
+      input [383:0] 	rz;
+
+      reg [383:0] 	px_shreg;
+      reg [383:0] 	py_shreg;
+      reg [383:0] 	pz_shreg;
+
+      reg [383:0] 	rx_shreg;
+      reg [383:0] 	ry_shreg;
+      reg [383:0] 	rz_shreg;
+
+      reg [383:0] 	q_shreg;
+
+      begin
+
+	 /* start filling memories */
+	 tb_xyzq_wren = 1;
+
+	 /* initialize shift registers */
+	 px_shreg = px;
+	 py_shreg = py;
+	 pz_shreg = pz;
+	 q_shreg  = Q;
+
+	 /* write all the words */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set addresses */
+	    tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* set data words */
+	    tb_px_data	= px_shreg[31:0];
+	    tb_py_data	= py_shreg[31:0];
+	    tb_pz_data	= pz_shreg[31:0];
+	    tb_q_data	= q_shreg[31:0];
+
+	    /* shift inputs */
+	    px_shreg = {{32{1'bX}}, px_shreg[383:32]};
+	    py_shreg = {{32{1'bX}}, py_shreg[383:32]};
+	    pz_shreg = {{32{1'bX}}, pz_shreg[383:32]};
+	    q_shreg  = {{32{1'bX}}, q_shreg[383:32]};
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	 end
+
+	 /* wipe addresses */
+	 tb_xyzq_addr = {WORD_COUNTER_WIDTH{1'bX}};
+
+	 /* wipe data words */
+	 tb_px_data = {32{1'bX}};
+	 tb_py_data = {32{1'bX}};
+	 tb_pz_data = {32{1'bX}};
+	 tb_q_data  = {32{1'bX}};
+
+	 /* stop filling memories */
+	 tb_xyzq_wren = 0;
+
+	 /* start operation */
+	 ena = 1;
+
+	 /* clear flag */
+	 #10 ena = 0;
+
+	 /* wait for operation to complete */
+	 while (!rdy) #10;
+
+	 /* read result */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set address */
+	    tb_xyzq_addr = w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	    /* store data word */
+	    rx_shreg = {tb_rx_data, rx_shreg[383:32]};
+	    ry_shreg = {tb_ry_data, ry_shreg[383:32]};
+	    rz_shreg = {tb_rz_data, rz_shreg[383:32]};
+
+	 end
+
+	 /* compare */
+	 t_ok =	(rx_shreg == rx) &&
+	       (ry_shreg == ry) &&
+	       (rz_shreg == rz);
+
+	 /* display results */
+	 $display("test_curve_doubler(): %s", t_ok ? "OK" : "ERROR");
+
+	 /* update global flag */
+	 ok = ok && t_ok;
+
+      end
+
+   endtask
+
+
 endmodule
 
 //------------------------------------------------------------------------------
diff --git a/bench/tb_curve_multiplier_384.v b/bench/tb_curve_multiplier_384.v
index c740f1f..cc02ee8 100644
--- a/bench/tb_curve_multiplier_384.v
+++ b/bench/tb_curve_multiplier_384.v
@@ -43,238 +43,238 @@
 module tb_curve_multiplier_384;
 
 
-		//
-		// Test Vectors
-		//
-	localparam	[383:0]	K_1	= 384'hc838b85253ef8dc7394fa5808a5183981c7deef5a69ba8f4f2117ffea39cfcd90e95f6cbc854abacab701d50c1f3cf24;
-	localparam	[383:0]	PX_1	= 384'h1fbac8eebd0cbf35640b39efe0808dd774debff20a2a329e91713baf7d7f3c3e81546d883730bee7e48678f857b02ca0;
-	localparam	[383:0]	PY_1	= 384'heb213103bd68ce343365a8a4c3d4555fa385f5330203bdd76ffad1f3affb95751c132007e1b240353cb0a4cf1693bdf9;
-
-	localparam	[383:0]	K_2	= 384'hdc6b44036989a196e39d1cdac000812f4bdd8b2db41bb33af51372585ebd1db63f0ce8275aa1fd45e2d2a735f8749359;
-	localparam	[383:0]	PX_2	= 384'ha0c27ec893092dea1e1bd2ccfed3cf945c8134ed0c9f81311a0f4a05942db8dbed8dd59f267471d5462aa14fe72de856;
-	localparam	[383:0]	PY_2	= 384'h855649409815bb91424eaca5fd76c97375d575d1422ec53d343bd33b847fdf0c11569685b528ab25493015428d7cf72b;
-
-	localparam	[383:0]	K_3	= 384'hffffffffffffffffffffffffffffffffffffffffffffffffc7634d81f4372ddf581a0db248b0a77aecec196accc52973;
-	localparam	[383:0]	PX_3	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
-	localparam	[383:0]	PY_3	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
-		
-		
-		//
-		// Core Parameters
-		//
-	localparam	WORD_COUNTER_WIDTH	=  4;
-	localparam	OPERAND_NUM_WORDS		= 12;
-	
-
-		//
-		// Clock (100 MHz)
-		//
-	reg clk = 1'b0;
-	always #5 clk = ~clk;
-
-	
-		//
-		// Inputs, Outputs
-		//
-	reg	rst_n;
-	reg	ena;
-	wire	rdy;
-
-	
-		//
-		// Buffers (K, PX, PY)
-		//
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_k_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_px_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_py_addr;
-	
-	wire										core_px_wren;
-	wire										core_py_wren;
-	
-	wire	[                32-1:0]	core_k_data;
-	wire	[                32-1:0]	core_px_data;
-	wire	[                32-1:0]	core_py_data;
-	
-	reg	[WORD_COUNTER_WIDTH-1:0]	tb_k_addr;
-	reg	[WORD_COUNTER_WIDTH-1:0]	tb_pxy_addr;
-	
-	reg										tb_k_wren;
-	
-	reg	[                  31:0]	tb_k_data;
-	wire	[                  31:0]	tb_px_data;
-	wire	[                  31:0]	tb_py_data;
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_k
-	(	.clk(clk),
-		.a_addr(tb_k_addr), .a_wr(tb_k_wren), .a_in(tb_k_data), .a_out(),
-		.b_addr(core_k_addr), .b_out(core_k_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_px
-	(	.clk(clk),
-		.a_addr(core_px_addr), .a_wr(core_px_wren), .a_in(core_px_data), .a_out(),
-		.b_addr(tb_pxy_addr), .b_out(tb_px_data)
-	);
-	
-	bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
-	bram_py
-	(	.clk(clk),
-		.a_addr(core_py_addr), .a_wr(core_py_wren), .a_in(core_py_data), .a_out(),
-		.b_addr(tb_pxy_addr), .b_out(tb_py_data)
-	);
-	
-	
-		//
-		// UUT
-		//
-	curve_mul_384 uut
-	(
-		.clk			(clk),
-		.rst_n		(rst_n),
-		
-		.ena			(ena),
-		.rdy			(rdy),
-		
-		.k_addr		(core_k_addr),
-		.rx_addr		(core_px_addr),
-		.ry_addr		(core_py_addr),
-		
-		.rx_wren		(core_px_wren),
-		.ry_wren		(core_py_wren),
-		
-		.k_din		(core_k_data),
-		
-		.rx_dout		(core_px_data),
-		.ry_dout		(core_py_data)
+   //
+   // Test Vectors
+   //
+   localparam	[383:0]	K_1	= 384'hc838b85253ef8dc7394fa5808a5183981c7deef5a69ba8f4f2117ffea39cfcd90e95f6cbc854abacab701d50c1f3cf24;
+   localparam	[383:0]	PX_1	= 384'h1fbac8eebd0cbf35640b39efe0808dd774debff20a2a329e91713baf7d7f3c3e81546d883730bee7e48678f857b02ca0;
+   localparam	[383:0]	PY_1	= 384'heb213103bd68ce343365a8a4c3d4555fa385f5330203bdd76ffad1f3affb95751c132007e1b240353cb0a4cf1693bdf9;
+
+   localparam	[383:0]	K_2	= 384'hdc6b44036989a196e39d1cdac000812f4bdd8b2db41bb33af51372585ebd1db63f0ce8275aa1fd45e2d2a735f8749359;
+   localparam	[383:0]	PX_2	= 384'ha0c27ec893092dea1e1bd2ccfed3cf945c8134ed0c9f81311a0f4a05942db8dbed8dd59f267471d5462aa14fe72de856;
+   localparam	[383:0]	PY_2	= 384'h855649409815bb91424eaca5fd76c97375d575d1422ec53d343bd33b847fdf0c11569685b528ab25493015428d7cf72b;
+
+   localparam	[383:0]	K_3	= 384'hffffffffffffffffffffffffffffffffffffffffffffffffc7634d81f4372ddf581a0db248b0a77aecec196accc52973;
+   localparam	[383:0]	PX_3	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
+   localparam	[383:0]	PY_3	= 384'h000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000;
+
+
+   //
+   // Core Parameters
+   //
+   localparam	WORD_COUNTER_WIDTH	=  4;
+   localparam	OPERAND_NUM_WORDS	= 12;
+
+
+   //
+   // Clock (100 MHz)
+   //
+   reg clk = 1'b0;
+   always #5 clk = ~clk;
+
+
+   //
+   // Inputs, Outputs
+   //
+   reg rst_n;
+   reg ena;
+   wire rdy;
+
+
+   //
+   // Buffers (K, PX, PY)
+   //
+   wire [WORD_COUNTER_WIDTH-1:0] core_k_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_px_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_py_addr;
+
+   wire 			 core_px_wren;
+   wire 			 core_py_wren;
+
+   wire [                32-1:0] core_k_data;
+   wire [                32-1:0] core_px_data;
+   wire [                32-1:0] core_py_data;
+
+   reg [WORD_COUNTER_WIDTH-1:0]  tb_k_addr;
+   reg [WORD_COUNTER_WIDTH-1:0]  tb_pxy_addr;
+
+   reg 				 tb_k_wren;
+
+   reg [                  31:0]  tb_k_data;
+   wire [                  31:0] tb_px_data;
+   wire [                  31:0] tb_py_data;
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_k
+     (	.clk(clk),
+	.a_addr(tb_k_addr), .a_wr(tb_k_wren), .a_in(tb_k_data), .a_out(),
+	.b_addr(core_k_addr), .b_out(core_k_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_px
+     (	.clk(clk),
+	.a_addr(core_px_addr), .a_wr(core_px_wren), .a_in(core_px_data), .a_out(),
+	.b_addr(tb_pxy_addr), .b_out(tb_px_data)
+	);
+
+   bram_1rw_1ro_readfirst # (.MEM_WIDTH(32), .MEM_ADDR_BITS(WORD_COUNTER_WIDTH))
+   bram_py
+     (	.clk(clk),
+	.a_addr(core_py_addr), .a_wr(core_py_wren), .a_in(core_py_data), .a_out(),
+	.b_addr(tb_pxy_addr), .b_out(tb_py_data)
 	);
-		
-		
-		//
-		// Testbench Routine
-		//
-	reg ok = 1;
-	initial begin
-		
-			/* initialize control inputs */
-		rst_n		= 0;
-		ena		= 0;
-		
-			/* wait for some time */
-		#200;
-		
-			/* de-assert reset */
-		rst_n		= 1;
-		
-			/* wait for some time */
-		#100;		
-		
-			/* run tests */
-		test_curve_multiplier(K_1, PX_1, PY_1);
-		test_curve_multiplier(K_2, PX_2, PY_2);
-		test_curve_multiplier(K_3, PX_3, PY_3);
-		
-			/* print result */
-		if (ok)	$display("tb_curve_multiplier_384: SUCCESS");
-		else		$display("tb_curve_multiplier_384: FAILURE");
-		//
-		//$finish;
-		//
-	end
-	
-	
-		//
-		// Test Task
-		//	
-	reg		p_ok;
-	
-	integer	w;
-
-	task test_curve_multiplier;
-	
-		input	[383:0]	k;
-		input	[383:0]	px;
-		input	[383:0]	py;
-		
-		reg	[383:0]	k_shreg;
-		reg	[383:0]	px_shreg;
-		reg	[383:0]	py_shreg;
-		
-		begin
-		
-				/* start filling memories */
-			tb_k_wren = 1;
-			
-				/* initialize shift registers */
-			k_shreg = k;
-			
-				/* write all the words */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set addresses */
-				tb_k_addr = w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* set data words */
-				tb_k_data	= k_shreg[31:0];
-				
-					/* shift inputs */
-				k_shreg = {{32{1'bX}}, k_shreg[383:32]};
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-			end
-			
-				/* wipe addresses */
-			tb_k_addr = {WORD_COUNTER_WIDTH{1'bX}};
-			
-				/* wipe data words */
-			tb_k_data = {32{1'bX}};
-			
-				/* stop filling memories */
-			tb_k_wren = 0;
-			
-				/* start operation */
-			ena = 1;
-			
-				/* clear flag */
-			#10 ena = 0;
-			
-				/* wait for operation to complete */
-			while (!rdy) #10;
-			
-				/* read result */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set address */
-				tb_pxy_addr = w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-					/* store data word */
-				px_shreg = {tb_px_data, px_shreg[383:32]};
-				py_shreg = {tb_py_data, py_shreg[383:32]};
-
-			end
-			
-				/* compare */
-			p_ok =	(px_shreg == px) &&
-						(py_shreg == py);
-
-				/* display results */
-			$display("test_curve_multiplier(): %s", p_ok ? "OK" : "ERROR");
-			
-				/* update global flag */
-			ok = ok && p_ok;
-		
-		end
-		
-	endtask
-	
+
+
+   //
+   // UUT
+   //
+   curve_mul_384 uut
+     (
+      .clk		(clk),
+      .rst_n		(rst_n),
+
+      .ena		(ena),
+      .rdy		(rdy),
+
+      .k_addr		(core_k_addr),
+      .rx_addr		(core_px_addr),
+      .ry_addr		(core_py_addr),
+
+      .rx_wren		(core_px_wren),
+      .ry_wren		(core_py_wren),
+
+      .k_din		(core_k_data),
+
+      .rx_dout		(core_px_data),
+      .ry_dout		(core_py_data)
+      );
+
+
+   //
+   // Testbench Routine
+   //
+   reg 			 ok = 1;
+   initial begin
+
+      /* initialize control inputs */
+      rst_n		= 0;
+      ena		= 0;
+
+      /* wait for some time */
+      #200;
+
+      /* de-assert reset */
+      rst_n		= 1;
+
+      /* wait for some time */
+      #100;
+
+      /* run tests */
+      test_curve_multiplier(K_1, PX_1, PY_1);
+      test_curve_multiplier(K_2, PX_2, PY_2);
+      test_curve_multiplier(K_3, PX_3, PY_3);
+
+      /* print result */
+      if (ok)	$display("tb_curve_multiplier_384: SUCCESS");
+      else	$display("tb_curve_multiplier_384: FAILURE");
+      //
+      //$finish;
+      //
+   end
+
+
+   //
+   // Test Task
+   //
+   reg		p_ok;
+
+   integer	w;
+
+   task test_curve_multiplier;
+
+      input [383:0]	k;
+      input [383:0] 	px;
+      input [383:0] 	py;
+
+      reg [383:0] 	k_shreg;
+      reg [383:0] 	px_shreg;
+      reg [383:0] 	py_shreg;
+
+      begin
+
+	 /* start filling memories */
+	 tb_k_wren = 1;
+
+	 /* initialize shift registers */
+	 k_shreg = k;
+
+	 /* write all the words */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set addresses */
+	    tb_k_addr = w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* set data words */
+	    tb_k_data	= k_shreg[31:0];
+
+	    /* shift inputs */
+	    k_shreg = {{32{1'bX}}, k_shreg[383:32]};
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	 end
+
+	 /* wipe addresses */
+	 tb_k_addr = {WORD_COUNTER_WIDTH{1'bX}};
+
+	 /* wipe data words */
+	 tb_k_data = {32{1'bX}};
+
+	 /* stop filling memories */
+	 tb_k_wren = 0;
+
+	 /* start operation */
+	 ena = 1;
+
+	 /* clear flag */
+	 #10 ena = 0;
+
+	 /* wait for operation to complete */
+	 while (!rdy) #10;
+
+	 /* read result */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set address */
+	    tb_pxy_addr = w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	    /* store data word */
+	    px_shreg = {tb_px_data, px_shreg[383:32]};
+	    py_shreg = {tb_py_data, py_shreg[383:32]};
+
+	 end
+
+	 /* compare */
+	 p_ok =	(px_shreg == px) &&
+	       (py_shreg == py);
+
+	 /* display results */
+	 $display("test_curve_multiplier(): %s", p_ok ? "OK" : "ERROR");
+
+	 /* update global flag */
+	 ok = ok && p_ok;
+
+      end
+
+   endtask
+
 endmodule
-
+
 
 //------------------------------------------------------------------------------
 // End-of-File
diff --git a/bench/tb_modular_multiplier_384.v b/bench/tb_modular_multiplier_384.v
index 13ffe5e..a5825d8 100644
--- a/bench/tb_modular_multiplier_384.v
+++ b/bench/tb_modular_multiplier_384.v
@@ -38,329 +38,329 @@
 
 //------------------------------------------------------------------------------
 `timescale 1ns / 1ps
-//------------------------------------------------------------------------------
-
-module tb_modular_multiplier_384;
-
-
-		//
-		// Test Vectors
-		//	
-	localparam	[383:0]	N		= 384'hfffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffeffffffff0000000000000000ffffffff;
-	
-	localparam	[383:0]	X_1	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7;
-	localparam	[383:0]	Y_1	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f;
-	localparam	[383:0]	P_1	= 384'h332e559389c970313cb29c4b55af5783821971a99c250daf84dc5d3cc441cb0a482e90de9d3ccd96b3c8c48b2ad3f025;
-	
-	localparam	[383:0]	X_2	= 384'haaf06bba82e9f590e29c71c219bea51723c5893ae8b0c8cf4c117c3efb57ab8d55fa1b428155ad278b5743911b13ea8a;
-	localparam	[383:0]	Y_2	= 384'hc9e821b569d9d390a26167406d6d23d6070be242d765eb831625ceec4a0f473ef59f4e30e2817e6285bce2846f15f19d;
-	localparam	[383:0]	P_2	= 384'haa1a9db70fba0a4c034777cdcd93e8bd6e9afa1171d43bdea0a16c32da20e7ebccb2fac9676f9d67a31e6f4f69e876e5;
-	
-	localparam	[383:0]	X_3	= 384'h1fbac8eebd0cbf35640b39efe0808dd774debff20a2a329e91713baf7d7f3c3e81546d883730bee7e48678f857b02ca0;
-	localparam	[383:0]	Y_3	= 384'heb213103bd68ce343365a8a4c3d4555fa385f5330203bdd76ffad1f3affb95751c132007e1b240353cb0a4cf1693bdf9;
-	localparam	[383:0]	P_3	= 384'h80f70000040a44b05f3752b7d5338f87e409b868f032911bda888451c13097039d66d9e7b0e3e799b9dd613d2524b7af;
-	
-	localparam	[383:0]	X_4	= 384'ha0c27ec893092dea1e1bd2ccfed3cf945c8134ed0c9f81311a0f4a05942db8dbed8dd59f267471d5462aa14fe72de856;
-	localparam	[383:0]	Y_4	= 384'h855649409815bb91424eaca5fd76c97375d575d1422ec53d343bd33b847fdf0c11569685b528ab25493015428d7cf72b;
-	localparam	[383:0]	P_4	= 384'h548e8456d5b3c36557a59914af514739a92908e59ddde731b8746891ad26199de955789e7cc34bfe966e3471c2684969;
-
-	
-		//
-		// Core Parameters
-		//
-	localparam	WORD_COUNTER_WIDTH	=  4;
-	localparam	OPERAND_NUM_WORDS		= 12;
-	
-
-		//
-		// Clock (100 MHz)
-		//
-	reg clk = 1'b0;
-	always #5 clk = ~clk;
-
-	
-		//
-		// Inputs, Outputs
-		//
-	reg	rst_n;
-	reg	ena;
-	wire	rdy;
-	
-	
-		//
-		// Buffers (X, Y, N, P)
-		//
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_x_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_y_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_n_addr;
-	wire	[WORD_COUNTER_WIDTH-1:0]	core_p_addr;
-	
-	wire										core_p_wren;
-	
-	wire	[                  31:0]	core_x_data;
-	wire	[                  31:0]	core_y_data;
-	wire	[                  31:0]	core_n_data;
-	wire	[                  31:0]	core_p_data;
-	
-	reg	[WORD_COUNTER_WIDTH-1:0]	tb_xyn_addr;
-	reg	[WORD_COUNTER_WIDTH-1:0]	tb_p_addr;
-	
-	reg										tb_xyn_wren;
-	
-	reg	[                  31:0]	tb_x_data;
-	reg	[                  31:0]	tb_y_data;
-	reg	[                  31:0]	tb_n_data;
-	wire	[                  31:0]	tb_p_data;
-	
-	bram_1rw_1ro_readfirst #
-	(
-		.MEM_WIDTH			(32),
-		.MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
-	)
-	bram_x
-	(
-		.clk		(clk),
-
-		.a_addr	(tb_xyn_addr),
-		.a_wr		(tb_xyn_wren),
-		.a_in		(tb_x_data),
-		.a_out	(),
-
-		.b_addr	(core_x_addr),
-		.b_out	(core_x_data)
-	);
-
-	bram_1rw_1ro_readfirst #
-	(
-		.MEM_WIDTH			(32),
-		.MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
-	)
-	bram_y
-	(
-		.clk		(clk),
-
-		.a_addr	(tb_xyn_addr),
-		.a_wr		(tb_xyn_wren),
-		.a_in		(tb_y_data),
-		.a_out	(),
-
-		.b_addr	(core_y_addr),
-		.b_out	(core_y_data)
-	);
-	
-	bram_1rw_1ro_readfirst #
-	(
-		.MEM_WIDTH			(32),
-		.MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
-	)
-	bram_n
-	(
-		.clk		(clk),
-
-		.a_addr	(tb_xyn_addr),
-		.a_wr		(tb_xyn_wren),
-		.a_in		(tb_n_data),
-		.a_out	(),
-
-		.b_addr	(core_n_addr),
-		.b_out	(core_n_data)
-	);
-	
-	bram_1rw_1ro_readfirst #
-	(
-		.MEM_WIDTH			(32),
-		.MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
-	)
-	bram_s
-	(
-		.clk		(clk),
-
-		.a_addr	(core_p_addr),
-		.a_wr		(core_p_wren),
-		.a_in		(core_p_data),
-		.a_out	(),
-
-		.b_addr	(tb_p_addr),
-		.b_out	(tb_p_data)
-	);
-	
-	
-		//
-		// UUT
-		//
-	modular_multiplier_384 uut
-	(
-		.clk		(clk),
-		.rst_n	(rst_n),
-		
-		.ena		(ena),
-		.rdy		(rdy),
-
-		.a_addr	(core_x_addr),
-		.b_addr	(core_y_addr),
-		.n_addr	(core_n_addr),
-		.p_addr	(core_p_addr),
-		.p_wren	(core_p_wren),
-		
-		.a_din	(core_x_data),
-		.b_din	(core_y_data),
-		.n_din	(core_n_data),
-		.p_dout	(core_p_data)
-	);
-
-		
-		//
-		// Testbench Routine
-		//
-	reg ok = 1;
-	initial begin
-		
-			/* initialize control inputs */
-		rst_n				= 0;
-		ena				= 0;
-		
-		tb_xyn_wren		= 0;
-		
-			/* wait for some time */
-		#200;
-		
-			/* de-assert reset */
-		rst_n				= 1;
-		
-			/* wait for some time */
-		#100;		
-		
-			/* run tests */
-		test_modular_multiplier_384(X_1, Y_1, N, P_1);
-		test_modular_multiplier_384(X_2, Y_2, N, P_2);
-		test_modular_multiplier_384(X_3, Y_3, N, P_3);
-		test_modular_multiplier_384(X_4, Y_4, N, P_4);
-		
-			/* print result */
-		if (ok)	$display("tb_modular_multiplier_384: SUCCESS");
-		else		$display("tb_modular_multiplier_384: FAILURE");
-		//
-		//$finish;
-		//
-	end
-	
-	
-		//
-		// Test Task
-		//
-	reg	[383:0]	p;
-	reg				p_ok;
-	
-	integer			w;
-	
-	reg	[767:0]	pp_full;
-	reg	[383:0]	pp_ref;
-	
-	task test_modular_multiplier_384;
-	
-		input	[383:0] x;
-		input	[383:0] y;
-		input	[383:0] n;
-		input	[383:0] pp;
-		
-		reg	[383:0]	x_shreg;
-		reg	[383:0]	y_shreg;
-		reg	[383:0]	n_shreg;
-		reg	[383:0]	p_shreg;
-	
-		begin
-		
-				/* start filling memories */
-			tb_xyn_wren	= 1;
-			
-				/* initialize shift registers */
-			x_shreg = x;
-			y_shreg = y;
-			n_shreg = n;
-			
-				/* write all the words */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set addresses */
-				tb_xyn_addr	= w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* set data words */
-				tb_x_data	= x_shreg[31:0];
-				tb_y_data	= y_shreg[31:0];
-				tb_n_data	= n_shreg[31:0];
-				
-					/* shift inputs */
-				x_shreg = {{32{1'bX}}, x_shreg[383:32]};
-				y_shreg = {{32{1'bX}}, y_shreg[383:32]};
-				n_shreg = {{32{1'bX}}, n_shreg[383:32]};
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-			end
-			
-				/* wipe addresses */
-			tb_xyn_addr	= {WORD_COUNTER_WIDTH{1'bX}};
-			
-				/* wipe data words */
-			tb_x_data	= {32{1'bX}};
-			tb_y_data	= {32{1'bX}};
-			tb_n_data	= {32{1'bX}};
-			
-				/* stop filling memories */
-			tb_xyn_wren	= 0;
-			
-				/* calculate reference value */
-			pp_full = {{384{1'b0}}, x} * {{384{1'b0}}, y};
-			pp_ref = pp_full % {{384{1'b0}}, n};
-			
-				/* compare reference value against hard-coded one */
-			if (pp_ref != pp) begin
-				$display("ERROR: pp_ref != pp");
-				$finish;
-			end
-			
-				/* start operation */
-			ena = 1;
-			
-				/* clear flag */
-			#10 ena = 0;
-			
-				/* wait for operation to complete */
-			while (!rdy) #10;
-			
-				/* read result */
-			for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
-				
-					/* set address */
-				tb_p_addr	= w[WORD_COUNTER_WIDTH-1:0];
-				
-					/* wait for 1 clock tick */
-				#10;
-				
-					/* store data word */
-				p_shreg = {tb_p_data, p_shreg[383:32]};
-
-			end				
-			
-				/* compare */
-			p_ok = (p_shreg == pp);
-
-				/* display results */
-			$display("test_modular_multiplier_384(): %s", p_ok ? "OK" : "ERROR");
-			
-				/* update flag */
-			ok = ok && p_ok;
-		
-		end
-		
-	endtask
-	
-
-	
-      
-endmodule
-
+//------------------------------------------------------------------------------
+
+module tb_modular_multiplier_384;
+
+
+   //
+   // Test Vectors
+   //
+   localparam	[383:0]	N	= 384'hfffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffeffffffff0000000000000000ffffffff;
+
+   localparam	[383:0]	X_1	= 384'haa87ca22be8b05378eb1c71ef320ad746e1d3b628ba79b9859f741e082542a385502f25dbf55296c3a545e3872760ab7;
+   localparam	[383:0]	Y_1	= 384'h3617de4a96262c6f5d9e98bf9292dc29f8f41dbd289a147ce9da3113b5f0b8c00a60b1ce1d7e819d7a431d7c90ea0e5f;
+   localparam	[383:0]	P_1	= 384'h332e559389c970313cb29c4b55af5783821971a99c250daf84dc5d3cc441cb0a482e90de9d3ccd96b3c8c48b2ad3f025;
+
+   localparam	[383:0]	X_2	= 384'haaf06bba82e9f590e29c71c219bea51723c5893ae8b0c8cf4c117c3efb57ab8d55fa1b428155ad278b5743911b13ea8a;
+   localparam	[383:0]	Y_2	= 384'hc9e821b569d9d390a26167406d6d23d6070be242d765eb831625ceec4a0f473ef59f4e30e2817e6285bce2846f15f19d;
+   localparam	[383:0]	P_2	= 384'haa1a9db70fba0a4c034777cdcd93e8bd6e9afa1171d43bdea0a16c32da20e7ebccb2fac9676f9d67a31e6f4f69e876e5;
+
+   localparam	[383:0]	X_3	= 384'h1fbac8eebd0cbf35640b39efe0808dd774debff20a2a329e91713baf7d7f3c3e81546d883730bee7e48678f857b02ca0;
+   localparam	[383:0]	Y_3	= 384'heb213103bd68ce343365a8a4c3d4555fa385f5330203bdd76ffad1f3affb95751c132007e1b240353cb0a4cf1693bdf9;
+   localparam	[383:0]	P_3	= 384'h80f70000040a44b05f3752b7d5338f87e409b868f032911bda888451c13097039d66d9e7b0e3e799b9dd613d2524b7af;
+
+   localparam	[383:0]	X_4	= 384'ha0c27ec893092dea1e1bd2ccfed3cf945c8134ed0c9f81311a0f4a05942db8dbed8dd59f267471d5462aa14fe72de856;
+   localparam	[383:0]	Y_4	= 384'h855649409815bb91424eaca5fd76c97375d575d1422ec53d343bd33b847fdf0c11569685b528ab25493015428d7cf72b;
+   localparam	[383:0]	P_4	= 384'h548e8456d5b3c36557a59914af514739a92908e59ddde731b8746891ad26199de955789e7cc34bfe966e3471c2684969;
+
+
+   //
+   // Core Parameters
+   //
+   localparam	WORD_COUNTER_WIDTH	=  4;
+   localparam	OPERAND_NUM_WORDS	= 12;
+
+
+   //
+   // Clock (100 MHz)
+   //
+   reg clk = 1'b0;
+   always #5 clk = ~clk;
+
+
+   //
+   // Inputs, Outputs
+   //
+   reg rst_n;
+   reg ena;
+   wire rdy;
+
+
+   //
+   // Buffers (X, Y, N, P)
+   //
+   wire [WORD_COUNTER_WIDTH-1:0] core_x_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_y_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_n_addr;
+   wire [WORD_COUNTER_WIDTH-1:0] core_p_addr;
+
+   wire 			 core_p_wren;
+
+   wire [                  31:0] core_x_data;
+   wire [                  31:0] core_y_data;
+   wire [                  31:0] core_n_data;
+   wire [                  31:0] core_p_data;
+
+   reg [WORD_COUNTER_WIDTH-1:0]  tb_xyn_addr;
+   reg [WORD_COUNTER_WIDTH-1:0]  tb_p_addr;
+
+   reg 				 tb_xyn_wren;
+
+   reg [                  31:0]  tb_x_data;
+   reg [                  31:0]  tb_y_data;
+   reg [                  31:0]  tb_n_data;
+   wire [                  31:0] tb_p_data;
+
+   bram_1rw_1ro_readfirst #
+     (
+      .MEM_WIDTH			(32),
+      .MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
+      )
+   bram_x
+     (
+      .clk		(clk),
+
+      .a_addr	(tb_xyn_addr),
+      .a_wr		(tb_xyn_wren),
+      .a_in		(tb_x_data),
+      .a_out	(),
+
+      .b_addr	(core_x_addr),
+      .b_out	(core_x_data)
+      );
+
+   bram_1rw_1ro_readfirst #
+     (
+      .MEM_WIDTH			(32),
+      .MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
+      )
+   bram_y
+     (
+      .clk		(clk),
+
+      .a_addr	(tb_xyn_addr),
+      .a_wr		(tb_xyn_wren),
+      .a_in		(tb_y_data),
+      .a_out	(),
+
+      .b_addr	(core_y_addr),
+      .b_out	(core_y_data)
+      );
+
+   bram_1rw_1ro_readfirst #
+     (
+      .MEM_WIDTH			(32),
+      .MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
+      )
+   bram_n
+     (
+      .clk		(clk),
+
+      .a_addr	(tb_xyn_addr),
+      .a_wr		(tb_xyn_wren),
+      .a_in		(tb_n_data),
+      .a_out	(),
+
+      .b_addr	(core_n_addr),
+      .b_out	(core_n_data)
+      );
+
+   bram_1rw_1ro_readfirst #
+     (
+      .MEM_WIDTH			(32),
+      .MEM_ADDR_BITS		(WORD_COUNTER_WIDTH)
+      )
+   bram_s
+     (
+      .clk	(clk),
+
+      .a_addr	(core_p_addr),
+      .a_wr	(core_p_wren),
+      .a_in	(core_p_data),
+      .a_out	(),
+
+      .b_addr	(tb_p_addr),
+      .b_out	(tb_p_data)
+      );
+
+
+   //
+   // UUT
+   //
+   modular_multiplier_384 uut
+     (
+      .clk	(clk),
+      .rst_n	(rst_n),
+
+      .ena	(ena),
+      .rdy	(rdy),
+
+      .a_addr	(core_x_addr),
+      .b_addr	(core_y_addr),
+      .n_addr	(core_n_addr),
+      .p_addr	(core_p_addr),
+      .p_wren	(core_p_wren),
+
+      .a_din	(core_x_data),
+      .b_din	(core_y_data),
+      .n_din	(core_n_data),
+      .p_dout	(core_p_data)
+      );
+
+
+   //
+   // Testbench Routine
+   //
+   reg			ok = 1;
+   initial begin
+
+      /* initialize control inputs */
+      rst_n		= 0;
+      ena		= 0;
+
+      tb_xyn_wren	= 0;
+
+      /* wait for some time */
+      #200;
+
+      /* de-assert reset */
+      rst_n		= 1;
+
+      /* wait for some time */
+      #100;
+
+      /* run tests */
+      test_modular_multiplier_384(X_1, Y_1, N, P_1);
+      test_modular_multiplier_384(X_2, Y_2, N, P_2);
+      test_modular_multiplier_384(X_3, Y_3, N, P_3);
+      test_modular_multiplier_384(X_4, Y_4, N, P_4);
+
+      /* print result */
+      if (ok)	$display("tb_modular_multiplier_384: SUCCESS");
+      else	$display("tb_modular_multiplier_384: FAILURE");
+      //
+      //$finish;
+      //
+   end
+
+
+   //
+   // Test Task
+   //
+   reg	[383:0]	p;
+   reg 		p_ok;
+
+   integer 	w;
+
+   reg [767:0] 	pp_full;
+   reg [383:0] 	pp_ref;
+
+   task test_modular_multiplier_384;
+
+      input	[383:0] x;
+      input [383:0] 	y;
+      input [383:0] 	n;
+      input [383:0] 	pp;
+
+      reg [383:0] 	x_shreg;
+      reg [383:0] 	y_shreg;
+      reg [383:0] 	n_shreg;
+      reg [383:0] 	p_shreg;
+
+      begin
+
+	 /* start filling memories */
+	 tb_xyn_wren	= 1;
+
+	 /* initialize shift registers */
+	 x_shreg = x;
+	 y_shreg = y;
+	 n_shreg = n;
+
+	 /* write all the words */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set addresses */
+	    tb_xyn_addr	= w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* set data words */
+	    tb_x_data	= x_shreg[31:0];
+	    tb_y_data	= y_shreg[31:0];
+	    tb_n_data	= n_shreg[31:0];
+
+	    /* shift inputs */
+	    x_shreg = {{32{1'bX}}, x_shreg[383:32]};
+	    y_shreg = {{32{1'bX}}, y_shreg[383:32]};
+	    n_shreg = {{32{1'bX}}, n_shreg[383:32]};
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	 end
+
+	 /* wipe addresses */
+	 tb_xyn_addr	= {WORD_COUNTER_WIDTH{1'bX}};
+
+	 /* wipe data words */
+	 tb_x_data	= {32{1'bX}};
+	 tb_y_data	= {32{1'bX}};
+	 tb_n_data	= {32{1'bX}};
+
+	 /* stop filling memories */
+	 tb_xyn_wren	= 0;
+
+	 /* calculate reference value */
+	 pp_full = {{384{1'b0}}, x} * {{384{1'b0}}, y};
+	 pp_ref = pp_full % {{384{1'b0}}, n};
+
+	 /* compare reference value against hard-coded one */
+	 if (pp_ref != pp) begin
+	    $display("ERROR: pp_ref != pp");
+	    $finish;
+	 end
+
+	 /* start operation */
+	 ena = 1;
+
+	 /* clear flag */
+	 #10 ena = 0;
+
+	 /* wait for operation to complete */
+	 while (!rdy) #10;
+
+	 /* read result */
+	 for (w=0; w<OPERAND_NUM_WORDS; w=w+1) begin
+
+	    /* set address */
+	    tb_p_addr	= w[WORD_COUNTER_WIDTH-1:0];
+
+	    /* wait for 1 clock tick */
+	    #10;
+
+	    /* store data word */
+	    p_shreg = {tb_p_data, p_shreg[383:32]};
+
+	 end
+
+	 /* compare */
+	 p_ok = (p_shreg == pp);
+
+	 /* display results */
+	 $display("test_modular_multiplier_384(): %s", p_ok ? "OK" : "ERROR");
+
+	 /* update flag */
+	 ok = ok && p_ok;
+
+      end
+
+   endtask
+
+
+
+
+endmodule
+
 //------------------------------------------------------------------------------
 // End-of-File
-//------------------------------------------------------------------------------
+//------------------------------------------------------------------------------
-- 
cgit v1.2.3