diff --git a/Argon2d/opt-sse/Makefile b/Argon2d/opt-sse/Makefile
index 1ff05e6..ca66726 100644
--- a/Argon2d/opt-sse/Makefile
+++ b/Argon2d/opt-sse/Makefile
@@ -1,4 +1,4 @@
-CFLAGS=-m64 -mavx -std=c++11 -pthread -O3
+CFLAGS=-m64 -mavx -std=c++0x -pthread -O3
 SOURCE=blake2b.cpp argon2d-opt-sse.cpp genkat.cpp
 CC=g++
 all:
diff --git a/Argon2d/opt-sse/argon2d-opt-sse.cpp b/Argon2d/opt-sse/argon2d-opt-sse.cpp
index eef7766..0910f3d 100644
--- a/Argon2d/opt-sse/argon2d-opt-sse.cpp
+++ b/Argon2d/opt-sse/argon2d-opt-sse.cpp
@@ -78,8 +78,13 @@ void free_memory(uint8_t **memory)
 }
 
 
+#define Swidth 8
+#define Swords (2 * (1 << Swidth))
+#define Sbytes (Swords * 8)
+#define Smask (((1 << Swidth) - 1) * 8)
+#define Smask2 (((uint64_t)Smask << 32) | Smask)
 
-void ComputeBlock(__m128i *state, uint8_t* ref_block_ptr, uint8_t* next_block_ptr)
+static void ComputeBlock(__m128i *state, uint8_t* ref_block_ptr, uint8_t* next_block_ptr, const uint64_t *Sboxes)
 {
 	__m128i ref_block[64];
 
@@ -101,55 +106,93 @@ void ComputeBlock(__m128i *state, uint8_t* ref_block_ptr, uint8_t* next_block_pt
 
 	for(unsigned i=0; i<1; ++i)
 {
+	const uint8_t *S0 = (const uint8_t *)Sboxes;
+	const uint8_t *S1 = (const uint8_t *)Sboxes + Sbytes / 2;
+	uint64_t x;
+
+	x = (uint32_t)_mm_cvtsi128_si32(state[0]) ^
+	    ((uint64_t)(uint32_t)_mm_cvtsi128_si32(state[63]) << 32);
+
+#define XFORM_ROUND { \
+	uint64_t xmask = x & Smask2; \
+	uint64_t s0 = *(const uint64_t *)(S0 + (uint32_t)xmask); \
+	uint64_t s1 = *(const uint64_t *)(S1 + (xmask >> 32)); \
+	x = (uint64_t)(x >> 32) * (uint32_t)x; \
+	x += s0; \
+	x ^= s1; \
+}
+
+#define XFORM_ROUNDS \
+	XFORM_ROUND XFORM_ROUND XFORM_ROUND XFORM_ROUND XFORM_ROUND XFORM_ROUND
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[0], state[1], state[2], state[3],
 		state[4], state[5], state[6], state[7]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[8], state[9], state[10], state[11],
 		state[12], state[13], state[14], state[15]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[16], state[17], state[18], state[19],
 		state[20], state[21], state[22], state[23]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[24], state[25], state[26], state[27],
 		state[28], state[29], state[30], state[31]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[32], state[33], state[34], state[35],
 		state[36], state[37], state[38], state[39]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[40], state[41], state[42], state[43],
 		state[44], state[45], state[46], state[47]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[48], state[49], state[50], state[51],
 		state[52], state[53], state[54], state[55]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[56], state[57], state[58], state[59],
 		state[60], state[61], state[62], state[63]);
 
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[0], state[8], state[16], state[24],
 		state[32], state[40], state[48], state[56]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[1], state[9], state[17], state[25],
 		state[33], state[41], state[49], state[57]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[2], state[10], state[18], state[26],
 		state[34], state[42], state[50], state[58])
+
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[3], state[11], state[19], state[27],
 		state[35], state[43], state[51], state[59]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[4], state[12], state[20], state[28],
 		state[36], state[44], state[52], state[60]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[5], state[13], state[21], state[29],
 		state[37], state[45], state[53], state[61]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[6], state[14], state[22], state[30],
 		state[38], state[46], state[54], state[62]);
 
+	XFORM_ROUNDS
 	BLAKE2_ROUND(state[7], state[15], state[23], state[31],
 		state[39], state[47], state[55], state[63]);
 
+	state[0] = _mm_add_epi32(state[0], _mm_cvtsi32_si128((uint32_t)x));
+	state[63] = _mm_add_epi32(state[63], _mm_cvtsi32_si128((uint32_t)(x >> 32)));
+
 	// BLAKE2 - end
 }
 	for (uint8_t i = 0; i< 64; i++)
@@ -234,6 +277,16 @@ void FillSegment(scheme_info_t *info, position_info_t pos)
 	uint32_t prev_block_offset; //offset of previous block
 	uint32_t prev_block_recalc=0; //number of the first block in the reference area in the previous slice 
 
+	uint64_t Sboxes[Swords];
+
+/* FIXME: The S-boxes must be {password, salt} dependent.
+ * We could use "memory", but it'd cause aliasing issues in ComputeBlock()
+ * for its writes to state[] vs. reads from S-boxes. */
+	{
+		for (uint32_t i = 0; i < Swords; i++)
+			Sboxes[i] = 6364136223846793005ULL * i;
+	}
+
 	if(0 == pos.pass && 0 == pos.slice) // First pass; first slice
 	{
 		start += 3;
@@ -252,7 +305,7 @@ void FillSegment(scheme_info_t *info, position_info_t pos)
 		uint32_t reference_block_offset = (pos.lane * segment_length) * BLOCK_SIZE;
 
 		// compute block
-		ComputeBlock(prev_block, memory+ reference_block_offset, memory+next_block_offset);//Computing third block in the segment
+		ComputeBlock(prev_block, memory+ reference_block_offset, memory+next_block_offset, Sboxes);//Computing third block in the segment
 
 		phi = _mm_extract_epi32(prev_block[0], 0);
 	}
@@ -311,7 +364,7 @@ void FillSegment(scheme_info_t *info, position_info_t pos)
 		}
 	
 		// compute block
-		ComputeBlock(prev_block, memory + reference_block_offset, memory+next_block_offset);
+		ComputeBlock(prev_block, memory + reference_block_offset, memory+next_block_offset, Sboxes);
 		phi = _mm_extract_epi32(prev_block[0], 0);
 		next_block_offset += BLOCK_SIZE;
 	}
@@ -473,6 +526,10 @@ int Argon2dOpt(uint8_t *out, uint32_t outlen, const uint8_t *msg, uint32_t msgle
 
 	Finalize(&info, out,outlen);
 
+#if 0
+	fwrite(memory, (size_t)m_cost<<10, 1, stderr);
+#endif
+
 	free_memory(&memory);
 
 #ifdef MEASURE
diff --git a/Argon2d/opt-sse/genkat.cpp b/Argon2d/opt-sse/genkat.cpp
index e16838e..cc0034a 100644
--- a/Argon2d/opt-sse/genkat.cpp
+++ b/Argon2d/opt-sse/genkat.cpp
@@ -134,7 +134,8 @@ void Run(void *out, uint32_t outlen, uint32_t inlen, uint32_t saltlen,
 	unsigned char one_array[256];
 	memset(one_array, 1, 256);
 
-	PHS(out, outlen, zero_array, inlen, one_array, saltlen, t_cost, m_cost);
+//	PHS(out, outlen, zero_array, inlen, one_array, saltlen, t_cost, m_cost);
+	Argon2dOpt((unsigned char *)out, outlen, zero_array, inlen, one_array, saltlen, NULL, 0, NULL, 0, t_cost, m_cost, thread_n);
 
 #ifdef _MEASURE
 	i3 = __rdtscp(&ui3);