diff --git a/src/Makefile.am b/src/Makefile.am
index e3a0f00..e98c623 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -25,11 +25,24 @@
 
 if HAVE_SSE3
 libosmocore_la_SOURCES += viterbi_sse.c
-# Per-object flags hack
-viterbi_sse.lo : CFLAGS += $(SIMD_FLAGS)
+if HAVE_SSE4_1
+viterbi_sse.lo : CFLAGS += -msse3 -msse4.1
+else
+viterbi_sse.lo : CFLAGS += -msse3
+endif
+
+if HAVE_AVX2
+libosmocore_la_SOURCES += viterbi_sse_avx.c
+if HAVE_SSE4_1
+viterbi_sse_avx.lo : CFLAGS += -msse3 -mavx2 -msse4.1
+else
+viterbi_sse_avx.lo : CFLAGS += -msse3 -mavx2
+endif
+endif
 endif
 
 BUILT_SOURCES = crc8gen.c crc16gen.c crc32gen.c crc64gen.c
+EXTRA_DIST = viterbi_sse_common.h
 
 if ENABLE_PLUGIN
 libosmocore_la_SOURCES += plugin.c
diff --git a/src/viterbi.c b/src/viterbi.c
index 2097a02..854754c 100644
--- a/src/viterbi.c
+++ b/src/viterbi.c
@@ -31,6 +31,18 @@
 #define BIT2NRZ(REG,N)	(((REG >> N) & 0x01) * 2 - 1) * -1
 #define NUM_STATES(K)	(K == 7 ? 64 : 16)
 
+#define INIT_POINTERS(simd) \
+{ \
+	osmo_conv_metrics_k5_n2 = osmo_conv_##simd##_metrics_k5_n2; \
+	osmo_conv_metrics_k5_n3 = osmo_conv_##simd##_metrics_k5_n3; \
+	osmo_conv_metrics_k5_n4 = osmo_conv_##simd##_metrics_k5_n4; \
+	osmo_conv_metrics_k7_n2 = osmo_conv_##simd##_metrics_k7_n2; \
+	osmo_conv_metrics_k7_n3 = osmo_conv_##simd##_metrics_k7_n3; \
+	osmo_conv_metrics_k7_n4 = osmo_conv_##simd##_metrics_k7_n4; \
+	vdec_malloc = &osmo_conv_##simd##_vdec_malloc; \
+	vdec_free = &osmo_conv_##simd##_vdec_free; \
+}
+
 static int init_complete = 0;
 
 __attribute__ ((visibility("hidden"))) int avx2_supported = 0;
@@ -38,19 +50,37 @@
 __attribute__ ((visibility("hidden"))) int sse41_supported = 0;
 
 /**
- * This pointers will be initialized by the osmo_conv_init()
- * depending on supported SIMD extensions.
+ * These pointers are being initialized at runtime by the
+ * osmo_conv_init() depending on supported SIMD extensions.
  */
 static int16_t *(*vdec_malloc)(size_t n);
 static void (*vdec_free)(int16_t *ptr);
 
-/* Forward malloc wrappers */
-int16_t *osmo_conv_vdec_malloc(size_t n);
-void osmo_conv_vdec_free(int16_t *ptr);
+void (*osmo_conv_metrics_k5_n2)(const int8_t *seq,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm);
+void (*osmo_conv_metrics_k5_n3)(const int8_t *seq,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm);
+void (*osmo_conv_metrics_k5_n4)(const int8_t *seq,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm);
+void (*osmo_conv_metrics_k7_n2)(const int8_t *seq,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm);
+void (*osmo_conv_metrics_k7_n3)(const int8_t *seq,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm);
+void (*osmo_conv_metrics_k7_n4)(const int8_t *seq,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm);
 
-#ifdef HAVE_SSE3
-int16_t *osmo_conv_vdec_malloc_sse3(size_t n);
-void osmo_conv_vdec_free_sse3(int16_t *ptr);
+/* Forward malloc wrappers */
+int16_t *osmo_conv_gen_vdec_malloc(size_t n);
+void osmo_conv_gen_vdec_free(int16_t *ptr);
+
+#if defined(HAVE_SSE3)
+int16_t *osmo_conv_sse_vdec_malloc(size_t n);
+void osmo_conv_sse_vdec_free(int16_t *ptr);
+#endif
+
+#if defined(HAVE_SSE3) && defined(HAVE_AVX2)
+int16_t *osmo_conv_sse_avx_vdec_malloc(size_t n);
+void osmo_conv_sse_avx_vdec_free(int16_t *ptr);
 #endif
 
 /* Forward Metric Units */
@@ -67,18 +97,33 @@
 void osmo_conv_gen_metrics_k7_n4(const int8_t *seq, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm);
 
-#ifdef HAVE_SSE3
-void osmo_conv_gen_metrics_k5_n2_sse(const int8_t *seq, const int16_t *out,
+#if defined(HAVE_SSE3)
+void osmo_conv_sse_metrics_k5_n2(const int8_t *seq, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm);
-void osmo_conv_gen_metrics_k5_n3_sse(const int8_t *seq, const int16_t *out,
+void osmo_conv_sse_metrics_k5_n3(const int8_t *seq, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm);
-void osmo_conv_gen_metrics_k5_n4_sse(const int8_t *seq, const int16_t *out,
+void osmo_conv_sse_metrics_k5_n4(const int8_t *seq, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm);
-void osmo_conv_gen_metrics_k7_n2_sse(const int8_t *seq, const int16_t *out,
+void osmo_conv_sse_metrics_k7_n2(const int8_t *seq, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm);
-void osmo_conv_gen_metrics_k7_n3_sse(const int8_t *seq, const int16_t *out,
+void osmo_conv_sse_metrics_k7_n3(const int8_t *seq, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm);
-void osmo_conv_gen_metrics_k7_n4_sse(const int8_t *seq, const int16_t *out,
+void osmo_conv_sse_metrics_k7_n4(const int8_t *seq, const int16_t *out,
+	int16_t *sums, int16_t *paths, int norm);
+#endif
+
+#if defined(HAVE_SSE3) && defined(HAVE_AVX2)
+void osmo_conv_sse_avx_metrics_k5_n2(const int8_t *seq, const int16_t *out,
+	int16_t *sums, int16_t *paths, int norm);
+void osmo_conv_sse_avx_metrics_k5_n3(const int8_t *seq, const int16_t *out,
+	int16_t *sums, int16_t *paths, int norm);
+void osmo_conv_sse_avx_metrics_k5_n4(const int8_t *seq, const int16_t *out,
+	int16_t *sums, int16_t *paths, int norm);
+void osmo_conv_sse_avx_metrics_k7_n2(const int8_t *seq, const int16_t *out,
+	int16_t *sums, int16_t *paths, int norm);
+void osmo_conv_sse_avx_metrics_k7_n3(const int8_t *seq, const int16_t *out,
+	int16_t *sums, int16_t *paths, int norm);
+void osmo_conv_sse_avx_metrics_k7_n4(const int8_t *seq, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm);
 #endif
 
@@ -488,31 +533,13 @@
 	if (dec->k == 5) {
 		switch (dec->n) {
 		case 2:
-		#ifdef HAVE_SSE3
-			dec->metric_func = !sse3_supported ?
-				osmo_conv_gen_metrics_k5_n2 :
-				osmo_conv_gen_metrics_k5_n2_sse;
-		#else
-			dec->metric_func = osmo_conv_gen_metrics_k5_n2;
-		#endif
+			dec->metric_func = osmo_conv_metrics_k5_n2;
 			break;
 		case 3:
-		#ifdef HAVE_SSE3
-			dec->metric_func = !sse3_supported ?
-				osmo_conv_gen_metrics_k5_n3 :
-				osmo_conv_gen_metrics_k5_n3_sse;
-		#else
-			dec->metric_func = osmo_conv_gen_metrics_k5_n3;
-		#endif
+			dec->metric_func = osmo_conv_metrics_k5_n3;
 			break;
 		case 4:
-		#ifdef HAVE_SSE3
-			dec->metric_func = !sse3_supported ?
-				osmo_conv_gen_metrics_k5_n4 :
-				osmo_conv_gen_metrics_k5_n4_sse;
-		#else
-			dec->metric_func = osmo_conv_gen_metrics_k5_n4;
-		#endif
+			dec->metric_func = osmo_conv_metrics_k5_n4;
 			break;
 		default:
 			goto fail;
@@ -520,31 +547,13 @@
 	} else if (dec->k == 7) {
 		switch (dec->n) {
 		case 2:
-		#ifdef HAVE_SSE3
-			dec->metric_func = !sse3_supported ?
-				osmo_conv_gen_metrics_k7_n2 :
-				osmo_conv_gen_metrics_k7_n2_sse;
-		#else
-			dec->metric_func = osmo_conv_gen_metrics_k7_n2;
-		#endif
+			dec->metric_func = osmo_conv_metrics_k7_n2;
 			break;
 		case 3:
-		#ifdef HAVE_SSE3
-			dec->metric_func = !sse3_supported ?
-				osmo_conv_gen_metrics_k7_n3 :
-				osmo_conv_gen_metrics_k7_n3_sse;
-		#else
-			dec->metric_func = osmo_conv_gen_metrics_k7_n3;
-		#endif
+			dec->metric_func = osmo_conv_metrics_k7_n3;
 			break;
 		case 4:
-		#ifdef HAVE_SSE3
-			dec->metric_func = !sse3_supported ?
-				osmo_conv_gen_metrics_k7_n4 :
-				osmo_conv_gen_metrics_k7_n4_sse;
-		#else
-			dec->metric_func = osmo_conv_gen_metrics_k7_n4;
-		#endif
+			dec->metric_func = osmo_conv_metrics_k7_n4;
 			break;
 		default:
 			goto fail;
@@ -656,14 +665,26 @@
 	#endif
 #endif
 
-#ifdef HAVE_SSE3
-	vdec_malloc = !sse3_supported ?
-		&osmo_conv_vdec_malloc : &osmo_conv_vdec_malloc_sse3;
-	vdec_free = !sse3_supported ?
-		&osmo_conv_vdec_free : &osmo_conv_vdec_free_sse3;
+/**
+ * Usage of curly braces is mandatory,
+ * because we use multi-line define.
+ */
+#if defined(HAVE_SSE3) && defined(HAVE_AVX2)
+	if (sse3_supported && avx2_supported) {
+		INIT_POINTERS(sse_avx);
+	} else if (sse3_supported) {
+		INIT_POINTERS(sse);
+	} else {
+		INIT_POINTERS(gen);
+	}
+#elif defined(HAVE_SSE3)
+	if (sse3_supported) {
+		INIT_POINTERS(sse);
+	} else {
+		INIT_POINTERS(gen);
+	}
 #else
-	vdec_malloc = &osmo_conv_vdec_malloc;
-	vdec_free = &osmo_conv_vdec_free;
+	INIT_POINTERS(gen);
 #endif
 }
 
diff --git a/src/viterbi_generic.c b/src/viterbi_generic.c
index 2ced615..6042668 100644
--- a/src/viterbi_generic.c
+++ b/src/viterbi_generic.c
@@ -129,13 +129,13 @@
 
 /* Not-aligned Memory Allocator */
 __attribute__ ((visibility("hidden")))
-int16_t *osmo_conv_vdec_malloc(size_t n)
+int16_t *osmo_conv_gen_vdec_malloc(size_t n)
 {
 	return (int16_t *) malloc(sizeof(int16_t) * n);
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_vdec_free(int16_t *ptr)
+void osmo_conv_gen_vdec_free(int16_t *ptr)
 {
 	free(ptr);
 }
diff --git a/src/viterbi_sse.c b/src/viterbi_sse.c
index 5ca21b0..f389085 100644
--- a/src/viterbi_sse.c
+++ b/src/viterbi_sse.c
@@ -21,159 +21,18 @@
  */
 
 #include <stdint.h>
+#include "config.h"
+
 #include <emmintrin.h>
 #include <tmmintrin.h>
 #include <xmmintrin.h>
 
-#include "config.h"
-
-#if defined(HAVE_SSE4_1) || defined(HAVE_SSE41)
-	#include <smmintrin.h>
-#endif
-
-#ifdef HAVE_AVX2
-	#include <immintrin.h>
+#if defined(HAVE_SSE4_1)
+#include <smmintrin.h>
 #endif
 
 #define SSE_ALIGN 16
 
-extern int sse41_supported;
-extern int sse3_supported;
-extern int avx2_supported;
-
-/* Octo-Viterbi butterfly
- * Compute 8-wide butterfly generating 16 path decisions and 16 accumulated
- * sums. Inputs all packed 16-bit integers in three 128-bit XMM registers.
- * Two intermediate registers are used and results are set in the upper 4
- * registers.
- *
- * Input:
- * M0 - Path metrics 0 (packed 16-bit integers)
- * M1 - Path metrics 1 (packed 16-bit integers)
- * M2 - Branch metrics (packed 16-bit integers)
- *
- * Output:
- * M2 - Selected and accumulated path metrics 0
- * M4 - Selected and accumulated path metrics 1
- * M3 - Path selections 0
- * M1 - Path selections 1
- */
-#define SSE_BUTTERFLY(M0, M1, M2, M3, M4) \
-{ \
-	M3 = _mm_adds_epi16(M0, M2); \
-	M4 = _mm_subs_epi16(M1, M2); \
-	M0 = _mm_subs_epi16(M0, M2); \
-	M1 = _mm_adds_epi16(M1, M2); \
-	M2 = _mm_max_epi16(M3, M4); \
-	M3 = _mm_or_si128(_mm_cmpgt_epi16(M3, M4), _mm_cmpeq_epi16(M3, M4)); \
-	M4 = _mm_max_epi16(M0, M1); \
-	M1 = _mm_or_si128(_mm_cmpgt_epi16(M0, M1), _mm_cmpeq_epi16(M0, M1)); \
-}
-
-/* Two lane deinterleaving K = 5:
- * Take 16 interleaved 16-bit integers and deinterleave to 2 packed 128-bit
- * registers. The operation summarized below. Four registers are used with
- * the lower 2 as input and upper 2 as output.
- *
- * In   - 10101010 10101010 10101010 10101010
- * Out  - 00000000 11111111 00000000 11111111
- *
- * Input:
- * M0:1 - Packed 16-bit integers
- *
- * Output:
- * M2:3 - Deinterleaved packed 16-bit integers
- */
-#define _I8_SHUFFLE_MASK 15, 14, 11, 10, 7, 6, 3, 2, 13, 12, 9, 8, 5, 4, 1, 0
-
-#define SSE_DEINTERLEAVE_K5(M0, M1, M2, M3) \
-{ \
-	M2 = _mm_set_epi8(_I8_SHUFFLE_MASK); \
-	M0 = _mm_shuffle_epi8(M0, M2); \
-	M1 = _mm_shuffle_epi8(M1, M2); \
-	M2 = _mm_unpacklo_epi64(M0, M1); \
-	M3 = _mm_unpackhi_epi64(M0, M1); \
-}
-
-/* Two lane deinterleaving K = 7:
- * Take 64 interleaved 16-bit integers and deinterleave to 8 packed 128-bit
- * registers. The operation summarized below. 16 registers are used with the
- * lower 8 as input and upper 8 as output.
- *
- * In   - 10101010 10101010 10101010 10101010 ...
- * Out  - 00000000 11111111 00000000 11111111 ...
- *
- * Input:
- * M0:7 - Packed 16-bit integers
- *
- * Output:
- * M8:15 - Deinterleaved packed 16-bit integers
- */
-#define SSE_DEINTERLEAVE_K7(M0, M1, M2, M3, M4, M5, M6, M7, \
-	M8, M9, M10, M11, M12, M13, M14, M15) \
-{ \
-	M8  = _mm_set_epi8(_I8_SHUFFLE_MASK); \
-	M0  = _mm_shuffle_epi8(M0, M8); \
-	M1  = _mm_shuffle_epi8(M1, M8); \
-	M2  = _mm_shuffle_epi8(M2, M8); \
-	M3  = _mm_shuffle_epi8(M3, M8); \
-	M4  = _mm_shuffle_epi8(M4, M8); \
-	M5  = _mm_shuffle_epi8(M5, M8); \
-	M6  = _mm_shuffle_epi8(M6, M8); \
-	M7  = _mm_shuffle_epi8(M7, M8); \
-	M8  = _mm_unpacklo_epi64(M0, M1); \
-	M9  = _mm_unpackhi_epi64(M0, M1); \
-	M10 = _mm_unpacklo_epi64(M2, M3); \
-	M11 = _mm_unpackhi_epi64(M2, M3); \
-	M12 = _mm_unpacklo_epi64(M4, M5); \
-	M13 = _mm_unpackhi_epi64(M4, M5); \
-	M14 = _mm_unpacklo_epi64(M6, M7); \
-	M15 = _mm_unpackhi_epi64(M6, M7); \
-}
-
-/* Generate branch metrics N = 2:
- * Compute 16 branch metrics from trellis outputs and input values.
- *
- * Input:
- * M0:3 - 16 x 2 packed 16-bit trellis outputs
- * M4   - Expanded and packed 16-bit input value
- *
- * Output:
- * M6:7 - 16 computed 16-bit branch metrics
- */
-#define SSE_BRANCH_METRIC_N2(M0, M1, M2, M3, M4, M6, M7) \
-{ \
-	M0 = _mm_sign_epi16(M4, M0); \
-	M1 = _mm_sign_epi16(M4, M1); \
-	M2 = _mm_sign_epi16(M4, M2); \
-	M3 = _mm_sign_epi16(M4, M3); \
-	M6 = _mm_hadds_epi16(M0, M1); \
-	M7 = _mm_hadds_epi16(M2, M3); \
-}
-
-/* Generate branch metrics N = 4:
- * Compute 8 branch metrics from trellis outputs and input values. This
- * macro is reused for N less than 4 where the extra soft input bits are
- * padded.
- *
- * Input:
- * M0:3 - 8 x 4 packed 16-bit trellis outputs
- * M4   - Expanded and packed 16-bit input value
- *
- * Output:
- * M5   - 8 computed 16-bit branch metrics
- */
-#define SSE_BRANCH_METRIC_N4(M0, M1, M2, M3, M4, M5) \
-{ \
-	M0 = _mm_sign_epi16(M4, M0); \
-	M1 = _mm_sign_epi16(M4, M1); \
-	M2 = _mm_sign_epi16(M4, M2); \
-	M3 = _mm_sign_epi16(M4, M3); \
-	M0 = _mm_hadds_epi16(M0, M1); \
-	M1 = _mm_hadds_epi16(M2, M3); \
-	M5 = _mm_hadds_epi16(M0, M1); \
-}
-
 /* Broadcast 16-bit integer
  * Repeat the low 16-bit integer to all elements of the 128-bit SSE
  * register. Only AVX2 has a dedicated broadcast instruction; use repeat
@@ -186,364 +45,17 @@
  * Output:
  * M0 - Contains broadcasted values
  */
-#ifdef HAVE_AVX2
-#define SSE_BROADCAST(M0) \
-{ \
-	if (avx2_supported) { \
-		M0 = _mm_broadcastw_epi16(M0); \
-	} else { \
-		M0 = _mm_unpacklo_epi16(M0, M0); \
-		M0 = _mm_unpacklo_epi32(M0, M0); \
-		M0 = _mm_unpacklo_epi64(M0, M0); \
-	} \
-}
-#else
 #define SSE_BROADCAST(M0) \
 { \
 	M0 = _mm_unpacklo_epi16(M0, M0); \
 	M0 = _mm_unpacklo_epi32(M0, M0); \
 	M0 = _mm_unpacklo_epi64(M0, M0); \
 }
-#endif
 
-/* Horizontal minimum
- * Compute horizontal minimum of packed unsigned 16-bit integers and place
- * result in the low 16-bit element of the source register. Only SSE 4.1
- * has a dedicated minpos instruction. One intermediate register is used
- * if SSE 4.1 is not available. This is a destructive operation and the
- * source register is overwritten.
- *
- * Input:
- * M0 - Packed unsigned 16-bit integers
- *
- * Output:
- * M0 - Minimum value placed in low 16-bit element
+/**
+ * Include common SSE implementation
  */
-#if defined(HAVE_SSE4_1) || defined(HAVE_SSE41)
-#define SSE_MINPOS(M0, M1) \
-{ \
-	if (sse41_supported) { \
-		M0 = _mm_minpos_epu16(M0); \
-	} else { \
-		M1 = _mm_shuffle_epi32(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
-		M0 = _mm_min_epi16(M0, M1); \
-		M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
-		M0 = _mm_min_epi16(M0, M1); \
-		M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 0, 1)); \
-		M0 = _mm_min_epi16(M0, M1); \
-	} \
-}
-#else
-#define SSE_MINPOS(M0, M1) \
-{ \
-	M1 = _mm_shuffle_epi32(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
-	M0 = _mm_min_epi16(M0, M1); \
-	M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
-	M0 = _mm_min_epi16(M0, M1); \
-	M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 0, 1)); \
-	M0 = _mm_min_epi16(M0, M1); \
-}
-#endif
-
-/* Normalize state metrics K = 5:
- * Compute 16-wide normalization by subtracting the smallest value from
- * all values. Inputs are 16 packed 16-bit integers across 2 XMM registers.
- * Two intermediate registers are used and normalized results are placed
- * in the originating locations.
- *
- * Input:
- * M0:1 - Path metrics 0:1 (packed 16-bit integers)
- *
- * Output:
- * M0:1 - Normalized path metrics 0:1
- */
-#define SSE_NORMALIZE_K5(M0, M1, M2, M3) \
-{ \
-	M2 = _mm_min_epi16(M0, M1); \
-	SSE_MINPOS(M2, M3) \
-	SSE_BROADCAST(M2) \
-	M0 = _mm_subs_epi16(M0, M2); \
-	M1 = _mm_subs_epi16(M1, M2); \
-}
-
-/* Normalize state metrics K = 7:
- * Compute 64-wide normalization by subtracting the smallest value from
- * all values. Inputs are 8 registers of accumulated sums and 4 temporary
- * registers. Normalized results are returned in the originating locations.
- *
- * Input:
- * M0:7 - Path metrics 0:7 (packed 16-bit integers)
- *
- * Output:
- * M0:7 - Normalized path metrics 0:7
- */
-#define SSE_NORMALIZE_K7(M0, M1, M2, M3, M4, M5, M6, M7, M8, M9, M10, M11) \
-{ \
-	M8  = _mm_min_epi16(M0, M1); \
-	M9  = _mm_min_epi16(M2, M3); \
-	M10 = _mm_min_epi16(M4, M5); \
-	M11 = _mm_min_epi16(M6, M7); \
-	M8  = _mm_min_epi16(M8, M9); \
-	M10 = _mm_min_epi16(M10, M11); \
-	M8  = _mm_min_epi16(M8, M10); \
-	SSE_MINPOS(M8, M9) \
-	SSE_BROADCAST(M8) \
-	M0  = _mm_subs_epi16(M0, M8); \
-	M1  = _mm_subs_epi16(M1, M8); \
-	M2  = _mm_subs_epi16(M2, M8); \
-	M3  = _mm_subs_epi16(M3, M8); \
-	M4  = _mm_subs_epi16(M4, M8); \
-	M5  = _mm_subs_epi16(M5, M8); \
-	M6  = _mm_subs_epi16(M6, M8); \
-	M7  = _mm_subs_epi16(M7, M8); \
-}
-
-/* Combined BMU/PMU (K=5, N=2)
- * Compute branch metrics followed by path metrics for half rate 16-state
- * trellis. 8 butterflies are computed. Accumulated path sums are not
- * preserved and read and written into the same memory location. Normalize
- * sums if requires.
- */
-__always_inline static void _sse_metrics_k5_n2(const int16_t *val,
-	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
-{
-	__m128i m0, m1, m2, m3, m4, m5, m6;
-
-	/* (BMU) Load input sequence */
-	m2 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
-
-	/* (BMU) Load trellis outputs */
-	m0 = _mm_load_si128((__m128i *) &out[0]);
-	m1 = _mm_load_si128((__m128i *) &out[8]);
-
-	/* (BMU) Compute branch metrics */
-	m0 = _mm_sign_epi16(m2, m0);
-	m1 = _mm_sign_epi16(m2, m1);
-	m2 = _mm_hadds_epi16(m0, m1);
-
-	/* (PMU) Load accumulated path metrics */
-	m0 = _mm_load_si128((__m128i *) &sums[0]);
-	m1 = _mm_load_si128((__m128i *) &sums[8]);
-
-	SSE_DEINTERLEAVE_K5(m0, m1, m3, m4)
-
-	/* (PMU) Butterflies: 0-7 */
-	SSE_BUTTERFLY(m3, m4, m2, m5, m6)
-
-	if (norm)
-		SSE_NORMALIZE_K5(m2, m6, m0, m1)
-
-	_mm_store_si128((__m128i *) &sums[0], m2);
-	_mm_store_si128((__m128i *) &sums[8], m6);
-	_mm_store_si128((__m128i *) &paths[0], m5);
-	_mm_store_si128((__m128i *) &paths[8], m4);
-}
-
-/* Combined BMU/PMU (K=5, N=3 and N=4)
- * Compute branch metrics followed by path metrics for 16-state and rates
- * to 1/4. 8 butterflies are computed. The input sequence is read four 16-bit
- * values at a time, and extra values should be set to zero for rates other
- * than 1/4. Normally only rates 1/3 and 1/4 are used as there is a
- * dedicated implementation of rate 1/2.
- */
-__always_inline static void _sse_metrics_k5_n4(const int16_t *val,
-	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
-{
-	__m128i m0, m1, m2, m3, m4, m5, m6;
-
-	/* (BMU) Load input sequence */
-	m4 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
-
-	/* (BMU) Load trellis outputs */
-	m0 = _mm_load_si128((__m128i *) &out[0]);
-	m1 = _mm_load_si128((__m128i *) &out[8]);
-	m2 = _mm_load_si128((__m128i *) &out[16]);
-	m3 = _mm_load_si128((__m128i *) &out[24]);
-
-	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m4, m2)
-
-	/* (PMU) Load accumulated path metrics */
-	m0 = _mm_load_si128((__m128i *) &sums[0]);
-	m1 = _mm_load_si128((__m128i *) &sums[8]);
-
-	SSE_DEINTERLEAVE_K5(m0, m1, m3, m4)
-
-	/* (PMU) Butterflies: 0-7 */
-	SSE_BUTTERFLY(m3, m4, m2, m5, m6)
-
-	if (norm)
-		SSE_NORMALIZE_K5(m2, m6, m0, m1)
-
-	_mm_store_si128((__m128i *) &sums[0], m2);
-	_mm_store_si128((__m128i *) &sums[8], m6);
-	_mm_store_si128((__m128i *) &paths[0], m5);
-	_mm_store_si128((__m128i *) &paths[8], m4);
-}
-
-/* Combined BMU/PMU (K=7, N=2)
- * Compute branch metrics followed by path metrics for half rate 64-state
- * trellis. 32 butterfly operations are computed. Deinterleaving path
- * metrics requires usage of the full SSE register file, so separate sums
- * before computing branch metrics to avoid register spilling.
- */
-__always_inline static void _sse_metrics_k7_n2(const int16_t *val,
-	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
-{
-	__m128i m0, m1, m2, m3, m4, m5, m6, m7, m8,
-		m9, m10, m11, m12, m13, m14, m15;
-
-	/* (PMU) Load accumulated path metrics */
-	m0 = _mm_load_si128((__m128i *) &sums[0]);
-	m1 = _mm_load_si128((__m128i *) &sums[8]);
-	m2 = _mm_load_si128((__m128i *) &sums[16]);
-	m3 = _mm_load_si128((__m128i *) &sums[24]);
-	m4 = _mm_load_si128((__m128i *) &sums[32]);
-	m5 = _mm_load_si128((__m128i *) &sums[40]);
-	m6 = _mm_load_si128((__m128i *) &sums[48]);
-	m7 = _mm_load_si128((__m128i *) &sums[56]);
-
-	/* (PMU) Deinterleave to even-odd registers */
-	SSE_DEINTERLEAVE_K7(m0, m1, m2, m3 ,m4 ,m5, m6, m7,
-			    m8, m9, m10, m11, m12, m13, m14, m15)
-
-	/* (BMU) Load input symbols */
-	m7 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
-
-	/* (BMU) Load trellis outputs */
-	m0 = _mm_load_si128((__m128i *) &out[0]);
-	m1 = _mm_load_si128((__m128i *) &out[8]);
-	m2 = _mm_load_si128((__m128i *) &out[16]);
-	m3 = _mm_load_si128((__m128i *) &out[24]);
-
-	SSE_BRANCH_METRIC_N2(m0, m1, m2, m3, m7, m4, m5)
-
-	m0 = _mm_load_si128((__m128i *) &out[32]);
-	m1 = _mm_load_si128((__m128i *) &out[40]);
-	m2 = _mm_load_si128((__m128i *) &out[48]);
-	m3 = _mm_load_si128((__m128i *) &out[56]);
-
-	SSE_BRANCH_METRIC_N2(m0, m1, m2, m3, m7, m6, m7)
-
-	/* (PMU) Butterflies: 0-15 */
-	SSE_BUTTERFLY(m8, m9, m4, m0, m1)
-	SSE_BUTTERFLY(m10, m11, m5, m2, m3)
-
-	_mm_store_si128((__m128i *) &paths[0], m0);
-	_mm_store_si128((__m128i *) &paths[8], m2);
-	_mm_store_si128((__m128i *) &paths[32], m9);
-	_mm_store_si128((__m128i *) &paths[40], m11);
-
-	/* (PMU) Butterflies: 17-31 */
-	SSE_BUTTERFLY(m12, m13, m6, m0, m2)
-	SSE_BUTTERFLY(m14, m15, m7, m9, m11)
-
-	_mm_store_si128((__m128i *) &paths[16], m0);
-	_mm_store_si128((__m128i *) &paths[24], m9);
-	_mm_store_si128((__m128i *) &paths[48], m13);
-	_mm_store_si128((__m128i *) &paths[56], m15);
-
-	if (norm)
-		SSE_NORMALIZE_K7(m4, m1, m5, m3, m6, m2,
-				 m7, m11, m0, m8, m9, m10)
-
-	_mm_store_si128((__m128i *) &sums[0], m4);
-	_mm_store_si128((__m128i *) &sums[8], m5);
-	_mm_store_si128((__m128i *) &sums[16], m6);
-	_mm_store_si128((__m128i *) &sums[24], m7);
-	_mm_store_si128((__m128i *) &sums[32], m1);
-	_mm_store_si128((__m128i *) &sums[40], m3);
-	_mm_store_si128((__m128i *) &sums[48], m2);
-	_mm_store_si128((__m128i *) &sums[56], m11);
-}
-
-/* Combined BMU/PMU (K=7, N=3 and N=4)
- * Compute branch metrics followed by path metrics for half rate 64-state
- * trellis. 32 butterfly operations are computed. Deinterleave path
- * metrics before computing branch metrics as in the half rate case.
- */
-__always_inline static void _sse_metrics_k7_n4(const int16_t *val,
-	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
-{
-	__m128i m0, m1, m2, m3, m4, m5, m6, m7;
-	__m128i m8, m9, m10, m11, m12, m13, m14, m15;
-
-	/* (PMU) Load accumulated path metrics */
-	m0 = _mm_load_si128((__m128i *) &sums[0]);
-	m1 = _mm_load_si128((__m128i *) &sums[8]);
-	m2 = _mm_load_si128((__m128i *) &sums[16]);
-	m3 = _mm_load_si128((__m128i *) &sums[24]);
-	m4 = _mm_load_si128((__m128i *) &sums[32]);
-	m5 = _mm_load_si128((__m128i *) &sums[40]);
-	m6 = _mm_load_si128((__m128i *) &sums[48]);
-	m7 = _mm_load_si128((__m128i *) &sums[56]);
-
-	/* (PMU) Deinterleave into even and odd packed registers */
-	SSE_DEINTERLEAVE_K7(m0, m1, m2, m3 ,m4 ,m5, m6, m7,
-			    m8, m9, m10, m11, m12, m13, m14, m15)
-
-	/* (BMU) Load and expand 8-bit input out to 16-bits */
-	m7 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
-
-	/* (BMU) Load and compute branch metrics */
-	m0 = _mm_load_si128((__m128i *) &out[0]);
-	m1 = _mm_load_si128((__m128i *) &out[8]);
-	m2 = _mm_load_si128((__m128i *) &out[16]);
-	m3 = _mm_load_si128((__m128i *) &out[24]);
-
-	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m4)
-
-	m0 = _mm_load_si128((__m128i *) &out[32]);
-	m1 = _mm_load_si128((__m128i *) &out[40]);
-	m2 = _mm_load_si128((__m128i *) &out[48]);
-	m3 = _mm_load_si128((__m128i *) &out[56]);
-
-	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m5)
-
-	m0 = _mm_load_si128((__m128i *) &out[64]);
-	m1 = _mm_load_si128((__m128i *) &out[72]);
-	m2 = _mm_load_si128((__m128i *) &out[80]);
-	m3 = _mm_load_si128((__m128i *) &out[88]);
-
-	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m6)
-
-	m0 = _mm_load_si128((__m128i *) &out[96]);
-	m1 = _mm_load_si128((__m128i *) &out[104]);
-	m2 = _mm_load_si128((__m128i *) &out[112]);
-	m3 = _mm_load_si128((__m128i *) &out[120]);
-
-	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m7)
-
-	/* (PMU) Butterflies: 0-15 */
-	SSE_BUTTERFLY(m8, m9, m4, m0, m1)
-	SSE_BUTTERFLY(m10, m11, m5, m2, m3)
-
-	_mm_store_si128((__m128i *) &paths[0], m0);
-	_mm_store_si128((__m128i *) &paths[8], m2);
-	_mm_store_si128((__m128i *) &paths[32], m9);
-	_mm_store_si128((__m128i *) &paths[40], m11);
-
-	/* (PMU) Butterflies: 17-31 */
-	SSE_BUTTERFLY(m12, m13, m6, m0, m2)
-	SSE_BUTTERFLY(m14, m15, m7, m9, m11)
-
-	_mm_store_si128((__m128i *) &paths[16], m0);
-	_mm_store_si128((__m128i *) &paths[24], m9);
-	_mm_store_si128((__m128i *) &paths[48], m13);
-	_mm_store_si128((__m128i *) &paths[56], m15);
-
-	if (norm)
-		SSE_NORMALIZE_K7(m4, m1, m5, m3, m6, m2,
-				 m7, m11, m0, m8, m9, m10)
-
-	_mm_store_si128((__m128i *) &sums[0], m4);
-	_mm_store_si128((__m128i *) &sums[8], m5);
-	_mm_store_si128((__m128i *) &sums[16], m6);
-	_mm_store_si128((__m128i *) &sums[24], m7);
-	_mm_store_si128((__m128i *) &sums[32], m1);
-	_mm_store_si128((__m128i *) &sums[40], m3);
-	_mm_store_si128((__m128i *) &sums[48], m2);
-	_mm_store_si128((__m128i *) &sums[56], m11);
-}
+#include <viterbi_sse_common.h>
 
 /* Aligned Memory Allocator
  * SSE requires 16-byte memory alignment. We store relevant trellis values
@@ -551,19 +63,19 @@
  * so the allocated memory is casted as such.
  */
 __attribute__ ((visibility("hidden")))
-int16_t *osmo_conv_vdec_malloc_sse3(size_t n)
+int16_t *osmo_conv_sse_vdec_malloc(size_t n)
 {
 	return (int16_t *) _mm_malloc(sizeof(int16_t) * n, SSE_ALIGN);
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_vdec_free_sse3(int16_t *ptr)
+void osmo_conv_sse_vdec_free(int16_t *ptr)
 {
 	_mm_free(ptr);
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_gen_metrics_k5_n2_sse(const int8_t *val, const int16_t *out,
+void osmo_conv_sse_metrics_k5_n2(const int8_t *val, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm)
 {
 	const int16_t _val[4] = { val[0], val[1], val[0], val[1] };
@@ -572,7 +84,7 @@
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_gen_metrics_k5_n3_sse(const int8_t *val, const int16_t *out,
+void osmo_conv_sse_metrics_k5_n3(const int8_t *val, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm)
 {
 	const int16_t _val[4] = { val[0], val[1], val[2], 0 };
@@ -581,7 +93,7 @@
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_gen_metrics_k5_n4_sse(const int8_t *val, const int16_t *out,
+void osmo_conv_sse_metrics_k5_n4(const int8_t *val, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm)
 {
 	const int16_t _val[4] = { val[0], val[1], val[2], val[3] };
@@ -590,7 +102,7 @@
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_gen_metrics_k7_n2_sse(const int8_t *val, const int16_t *out,
+void osmo_conv_sse_metrics_k7_n2(const int8_t *val, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm)
 {
 	const int16_t _val[4] = { val[0], val[1], val[0], val[1] };
@@ -599,7 +111,7 @@
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_gen_metrics_k7_n3_sse(const int8_t *val, const int16_t *out,
+void osmo_conv_sse_metrics_k7_n3(const int8_t *val, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm)
 {
 	const int16_t _val[4] = { val[0], val[1], val[2], 0 };
@@ -608,7 +120,7 @@
 }
 
 __attribute__ ((visibility("hidden")))
-void osmo_conv_gen_metrics_k7_n4_sse(const int8_t *val, const int16_t *out,
+void osmo_conv_sse_metrics_k7_n4(const int8_t *val, const int16_t *out,
 	int16_t *sums, int16_t *paths, int norm)
 {
 	const int16_t _val[4] = { val[0], val[1], val[2], val[3] };
diff --git a/src/viterbi_sse_avx.c b/src/viterbi_sse_avx.c
new file mode 100644
index 0000000..b4c45a6
--- /dev/null
+++ b/src/viterbi_sse_avx.c
@@ -0,0 +1,129 @@
+/*
+ * Intel SSE + AVX Viterbi decoder
+ *
+ * Copyright (C) 2013, 2014 Thomas Tsou <tom@tsou.cc>
+ *
+ * All Rights Reserved
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#include <stdint.h>
+#include "config.h"
+
+#include <emmintrin.h>
+#include <tmmintrin.h>
+#include <xmmintrin.h>
+#include <immintrin.h>
+
+#if defined(HAVE_SSE4_1)
+#include <smmintrin.h>
+#endif
+
+#define SSE_ALIGN 16
+
+
+/* Broadcast 16-bit integer
+ * Repeat the low 16-bit integer to all elements of the 128-bit SSE
+ * register. Only AVX2 has a dedicated broadcast instruction; use repeat
+ * unpacks for SSE only architectures. This is a destructive operation and
+ * the source register is overwritten.
+ *
+ * Input:
+ * M0 - Low 16-bit element is read
+ *
+ * Output:
+ * M0 - Contains broadcasted values
+ */
+#define SSE_BROADCAST(M0) \
+{ \
+	M0 = _mm_broadcastw_epi16(M0); \
+}
+
+/**
+ * Include common SSE implementation
+ */
+#include <viterbi_sse_common.h>
+
+/* Aligned Memory Allocator
+ * SSE requires 16-byte memory alignment. We store relevant trellis values
+ * (accumulated sums, outputs, and path decisions) as 16 bit signed integers
+ * so the allocated memory is casted as such.
+ */
+__attribute__ ((visibility("hidden")))
+int16_t *osmo_conv_sse_avx_vdec_malloc(size_t n)
+{
+	return (int16_t *) _mm_malloc(sizeof(int16_t) * n, SSE_ALIGN);
+}
+
+__attribute__ ((visibility("hidden")))
+void osmo_conv_sse_avx_vdec_free(int16_t *ptr)
+{
+	_mm_free(ptr);
+}
+
+__attribute__ ((visibility("hidden")))
+void osmo_conv_sse_avx_metrics_k5_n2(const int8_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	const int16_t _val[4] = { val[0], val[1], val[0], val[1] };
+
+	_sse_metrics_k5_n2(_val, out, sums, paths, norm);
+}
+
+__attribute__ ((visibility("hidden")))
+void osmo_conv_sse_avx_metrics_k5_n3(const int8_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	const int16_t _val[4] = { val[0], val[1], val[2], 0 };
+
+	_sse_metrics_k5_n4(_val, out, sums, paths, norm);
+}
+
+__attribute__ ((visibility("hidden")))
+void osmo_conv_sse_avx_metrics_k5_n4(const int8_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	const int16_t _val[4] = { val[0], val[1], val[2], val[3] };
+
+	_sse_metrics_k5_n4(_val, out, sums, paths, norm);
+}
+
+__attribute__ ((visibility("hidden")))
+void osmo_conv_sse_avx_metrics_k7_n2(const int8_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	const int16_t _val[4] = { val[0], val[1], val[0], val[1] };
+
+	_sse_metrics_k7_n2(_val, out, sums, paths, norm);
+}
+
+__attribute__ ((visibility("hidden")))
+void osmo_conv_sse_avx_metrics_k7_n3(const int8_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	const int16_t _val[4] = { val[0], val[1], val[2], 0 };
+
+	_sse_metrics_k7_n4(_val, out, sums, paths, norm);
+}
+
+__attribute__ ((visibility("hidden")))
+void osmo_conv_sse_avx_metrics_k7_n4(const int8_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	const int16_t _val[4] = { val[0], val[1], val[2], val[3] };
+
+	_sse_metrics_k7_n4(_val, out, sums, paths, norm);
+}
diff --git a/src/viterbi_sse_common.h b/src/viterbi_sse_common.h
new file mode 100644
index 0000000..7d48c94
--- /dev/null
+++ b/src/viterbi_sse_common.h
@@ -0,0 +1,495 @@
+/*
+ * Intel SSE Viterbi decoder
+ *
+ * Copyright (C) 2013, 2014 Thomas Tsou <tom@tsou.cc>
+ *
+ * All Rights Reserved
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+extern int sse41_supported;
+
+/* Octo-Viterbi butterfly
+ * Compute 8-wide butterfly generating 16 path decisions and 16 accumulated
+ * sums. Inputs all packed 16-bit integers in three 128-bit XMM registers.
+ * Two intermediate registers are used and results are set in the upper 4
+ * registers.
+ *
+ * Input:
+ * M0 - Path metrics 0 (packed 16-bit integers)
+ * M1 - Path metrics 1 (packed 16-bit integers)
+ * M2 - Branch metrics (packed 16-bit integers)
+ *
+ * Output:
+ * M2 - Selected and accumulated path metrics 0
+ * M4 - Selected and accumulated path metrics 1
+ * M3 - Path selections 0
+ * M1 - Path selections 1
+ */
+#define SSE_BUTTERFLY(M0, M1, M2, M3, M4) \
+{ \
+	M3 = _mm_adds_epi16(M0, M2); \
+	M4 = _mm_subs_epi16(M1, M2); \
+	M0 = _mm_subs_epi16(M0, M2); \
+	M1 = _mm_adds_epi16(M1, M2); \
+	M2 = _mm_max_epi16(M3, M4); \
+	M3 = _mm_or_si128(_mm_cmpgt_epi16(M3, M4), _mm_cmpeq_epi16(M3, M4)); \
+	M4 = _mm_max_epi16(M0, M1); \
+	M1 = _mm_or_si128(_mm_cmpgt_epi16(M0, M1), _mm_cmpeq_epi16(M0, M1)); \
+}
+
+/* Two lane deinterleaving K = 5:
+ * Take 16 interleaved 16-bit integers and deinterleave to 2 packed 128-bit
+ * registers. The operation summarized below. Four registers are used with
+ * the lower 2 as input and upper 2 as output.
+ *
+ * In   - 10101010 10101010 10101010 10101010
+ * Out  - 00000000 11111111 00000000 11111111
+ *
+ * Input:
+ * M0:1 - Packed 16-bit integers
+ *
+ * Output:
+ * M2:3 - Deinterleaved packed 16-bit integers
+ */
+#define _I8_SHUFFLE_MASK 15, 14, 11, 10, 7, 6, 3, 2, 13, 12, 9, 8, 5, 4, 1, 0
+
+#define SSE_DEINTERLEAVE_K5(M0, M1, M2, M3) \
+{ \
+	M2 = _mm_set_epi8(_I8_SHUFFLE_MASK); \
+	M0 = _mm_shuffle_epi8(M0, M2); \
+	M1 = _mm_shuffle_epi8(M1, M2); \
+	M2 = _mm_unpacklo_epi64(M0, M1); \
+	M3 = _mm_unpackhi_epi64(M0, M1); \
+}
+
+/* Two lane deinterleaving K = 7:
+ * Take 64 interleaved 16-bit integers and deinterleave to 8 packed 128-bit
+ * registers. The operation summarized below. 16 registers are used with the
+ * lower 8 as input and upper 8 as output.
+ *
+ * In   - 10101010 10101010 10101010 10101010 ...
+ * Out  - 00000000 11111111 00000000 11111111 ...
+ *
+ * Input:
+ * M0:7 - Packed 16-bit integers
+ *
+ * Output:
+ * M8:15 - Deinterleaved packed 16-bit integers
+ */
+#define SSE_DEINTERLEAVE_K7(M0, M1, M2, M3, M4, M5, M6, M7, \
+	M8, M9, M10, M11, M12, M13, M14, M15) \
+{ \
+	M8  = _mm_set_epi8(_I8_SHUFFLE_MASK); \
+	M0  = _mm_shuffle_epi8(M0, M8); \
+	M1  = _mm_shuffle_epi8(M1, M8); \
+	M2  = _mm_shuffle_epi8(M2, M8); \
+	M3  = _mm_shuffle_epi8(M3, M8); \
+	M4  = _mm_shuffle_epi8(M4, M8); \
+	M5  = _mm_shuffle_epi8(M5, M8); \
+	M6  = _mm_shuffle_epi8(M6, M8); \
+	M7  = _mm_shuffle_epi8(M7, M8); \
+	M8  = _mm_unpacklo_epi64(M0, M1); \
+	M9  = _mm_unpackhi_epi64(M0, M1); \
+	M10 = _mm_unpacklo_epi64(M2, M3); \
+	M11 = _mm_unpackhi_epi64(M2, M3); \
+	M12 = _mm_unpacklo_epi64(M4, M5); \
+	M13 = _mm_unpackhi_epi64(M4, M5); \
+	M14 = _mm_unpacklo_epi64(M6, M7); \
+	M15 = _mm_unpackhi_epi64(M6, M7); \
+}
+
+/* Generate branch metrics N = 2:
+ * Compute 16 branch metrics from trellis outputs and input values.
+ *
+ * Input:
+ * M0:3 - 16 x 2 packed 16-bit trellis outputs
+ * M4   - Expanded and packed 16-bit input value
+ *
+ * Output:
+ * M6:7 - 16 computed 16-bit branch metrics
+ */
+#define SSE_BRANCH_METRIC_N2(M0, M1, M2, M3, M4, M6, M7) \
+{ \
+	M0 = _mm_sign_epi16(M4, M0); \
+	M1 = _mm_sign_epi16(M4, M1); \
+	M2 = _mm_sign_epi16(M4, M2); \
+	M3 = _mm_sign_epi16(M4, M3); \
+	M6 = _mm_hadds_epi16(M0, M1); \
+	M7 = _mm_hadds_epi16(M2, M3); \
+}
+
+/* Generate branch metrics N = 4:
+ * Compute 8 branch metrics from trellis outputs and input values. This
+ * macro is reused for N less than 4 where the extra soft input bits are
+ * padded.
+ *
+ * Input:
+ * M0:3 - 8 x 4 packed 16-bit trellis outputs
+ * M4   - Expanded and packed 16-bit input value
+ *
+ * Output:
+ * M5   - 8 computed 16-bit branch metrics
+ */
+#define SSE_BRANCH_METRIC_N4(M0, M1, M2, M3, M4, M5) \
+{ \
+	M0 = _mm_sign_epi16(M4, M0); \
+	M1 = _mm_sign_epi16(M4, M1); \
+	M2 = _mm_sign_epi16(M4, M2); \
+	M3 = _mm_sign_epi16(M4, M3); \
+	M0 = _mm_hadds_epi16(M0, M1); \
+	M1 = _mm_hadds_epi16(M2, M3); \
+	M5 = _mm_hadds_epi16(M0, M1); \
+}
+
+/* Horizontal minimum
+ * Compute horizontal minimum of packed unsigned 16-bit integers and place
+ * result in the low 16-bit element of the source register. Only SSE 4.1
+ * has a dedicated minpos instruction. One intermediate register is used
+ * if SSE 4.1 is not available. This is a destructive operation and the
+ * source register is overwritten.
+ *
+ * Input:
+ * M0 - Packed unsigned 16-bit integers
+ *
+ * Output:
+ * M0 - Minimum value placed in low 16-bit element
+ */
+#if defined(HAVE_SSE4_1) || defined(HAVE_SSE41)
+#define SSE_MINPOS(M0, M1) \
+{ \
+	if (sse41_supported) { \
+		M0 = _mm_minpos_epu16(M0); \
+	} else { \
+		M1 = _mm_shuffle_epi32(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
+		M0 = _mm_min_epi16(M0, M1); \
+		M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
+		M0 = _mm_min_epi16(M0, M1); \
+		M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 0, 1)); \
+		M0 = _mm_min_epi16(M0, M1); \
+	} \
+}
+#else
+#define SSE_MINPOS(M0, M1) \
+{ \
+	M1 = _mm_shuffle_epi32(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
+	M0 = _mm_min_epi16(M0, M1); \
+	M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 3, 2)); \
+	M0 = _mm_min_epi16(M0, M1); \
+	M1 = _mm_shufflelo_epi16(M0, _MM_SHUFFLE(0, 0, 0, 1)); \
+	M0 = _mm_min_epi16(M0, M1); \
+}
+#endif
+
+/* Normalize state metrics K = 5:
+ * Compute 16-wide normalization by subtracting the smallest value from
+ * all values. Inputs are 16 packed 16-bit integers across 2 XMM registers.
+ * Two intermediate registers are used and normalized results are placed
+ * in the originating locations.
+ *
+ * Input:
+ * M0:1 - Path metrics 0:1 (packed 16-bit integers)
+ *
+ * Output:
+ * M0:1 - Normalized path metrics 0:1
+ */
+#define SSE_NORMALIZE_K5(M0, M1, M2, M3) \
+{ \
+	M2 = _mm_min_epi16(M0, M1); \
+	SSE_MINPOS(M2, M3) \
+	SSE_BROADCAST(M2) \
+	M0 = _mm_subs_epi16(M0, M2); \
+	M1 = _mm_subs_epi16(M1, M2); \
+}
+
+/* Normalize state metrics K = 7:
+ * Compute 64-wide normalization by subtracting the smallest value from
+ * all values. Inputs are 8 registers of accumulated sums and 4 temporary
+ * registers. Normalized results are returned in the originating locations.
+ *
+ * Input:
+ * M0:7 - Path metrics 0:7 (packed 16-bit integers)
+ *
+ * Output:
+ * M0:7 - Normalized path metrics 0:7
+ */
+#define SSE_NORMALIZE_K7(M0, M1, M2, M3, M4, M5, M6, M7, M8, M9, M10, M11) \
+{ \
+	M8  = _mm_min_epi16(M0, M1); \
+	M9  = _mm_min_epi16(M2, M3); \
+	M10 = _mm_min_epi16(M4, M5); \
+	M11 = _mm_min_epi16(M6, M7); \
+	M8  = _mm_min_epi16(M8, M9); \
+	M10 = _mm_min_epi16(M10, M11); \
+	M8  = _mm_min_epi16(M8, M10); \
+	SSE_MINPOS(M8, M9) \
+	SSE_BROADCAST(M8) \
+	M0  = _mm_subs_epi16(M0, M8); \
+	M1  = _mm_subs_epi16(M1, M8); \
+	M2  = _mm_subs_epi16(M2, M8); \
+	M3  = _mm_subs_epi16(M3, M8); \
+	M4  = _mm_subs_epi16(M4, M8); \
+	M5  = _mm_subs_epi16(M5, M8); \
+	M6  = _mm_subs_epi16(M6, M8); \
+	M7  = _mm_subs_epi16(M7, M8); \
+}
+
+/* Combined BMU/PMU (K=5, N=2)
+ * Compute branch metrics followed by path metrics for half rate 16-state
+ * trellis. 8 butterflies are computed. Accumulated path sums are not
+ * preserved and read and written into the same memory location. Normalize
+ * sums if requires.
+ */
+__always_inline static void _sse_metrics_k5_n2(const int16_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	__m128i m0, m1, m2, m3, m4, m5, m6;
+
+	/* (BMU) Load input sequence */
+	m2 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
+
+	/* (BMU) Load trellis outputs */
+	m0 = _mm_load_si128((__m128i *) &out[0]);
+	m1 = _mm_load_si128((__m128i *) &out[8]);
+
+	/* (BMU) Compute branch metrics */
+	m0 = _mm_sign_epi16(m2, m0);
+	m1 = _mm_sign_epi16(m2, m1);
+	m2 = _mm_hadds_epi16(m0, m1);
+
+	/* (PMU) Load accumulated path metrics */
+	m0 = _mm_load_si128((__m128i *) &sums[0]);
+	m1 = _mm_load_si128((__m128i *) &sums[8]);
+
+	SSE_DEINTERLEAVE_K5(m0, m1, m3, m4)
+
+	/* (PMU) Butterflies: 0-7 */
+	SSE_BUTTERFLY(m3, m4, m2, m5, m6)
+
+	if (norm)
+		SSE_NORMALIZE_K5(m2, m6, m0, m1)
+
+	_mm_store_si128((__m128i *) &sums[0], m2);
+	_mm_store_si128((__m128i *) &sums[8], m6);
+	_mm_store_si128((__m128i *) &paths[0], m5);
+	_mm_store_si128((__m128i *) &paths[8], m4);
+}
+
+/* Combined BMU/PMU (K=5, N=3 and N=4)
+ * Compute branch metrics followed by path metrics for 16-state and rates
+ * to 1/4. 8 butterflies are computed. The input sequence is read four 16-bit
+ * values at a time, and extra values should be set to zero for rates other
+ * than 1/4. Normally only rates 1/3 and 1/4 are used as there is a
+ * dedicated implementation of rate 1/2.
+ */
+__always_inline static void _sse_metrics_k5_n4(const int16_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	__m128i m0, m1, m2, m3, m4, m5, m6;
+
+	/* (BMU) Load input sequence */
+	m4 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
+
+	/* (BMU) Load trellis outputs */
+	m0 = _mm_load_si128((__m128i *) &out[0]);
+	m1 = _mm_load_si128((__m128i *) &out[8]);
+	m2 = _mm_load_si128((__m128i *) &out[16]);
+	m3 = _mm_load_si128((__m128i *) &out[24]);
+
+	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m4, m2)
+
+	/* (PMU) Load accumulated path metrics */
+	m0 = _mm_load_si128((__m128i *) &sums[0]);
+	m1 = _mm_load_si128((__m128i *) &sums[8]);
+
+	SSE_DEINTERLEAVE_K5(m0, m1, m3, m4)
+
+	/* (PMU) Butterflies: 0-7 */
+	SSE_BUTTERFLY(m3, m4, m2, m5, m6)
+
+	if (norm)
+		SSE_NORMALIZE_K5(m2, m6, m0, m1)
+
+	_mm_store_si128((__m128i *) &sums[0], m2);
+	_mm_store_si128((__m128i *) &sums[8], m6);
+	_mm_store_si128((__m128i *) &paths[0], m5);
+	_mm_store_si128((__m128i *) &paths[8], m4);
+}
+
+/* Combined BMU/PMU (K=7, N=2)
+ * Compute branch metrics followed by path metrics for half rate 64-state
+ * trellis. 32 butterfly operations are computed. Deinterleaving path
+ * metrics requires usage of the full SSE register file, so separate sums
+ * before computing branch metrics to avoid register spilling.
+ */
+__always_inline static void _sse_metrics_k7_n2(const int16_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	__m128i m0, m1, m2, m3, m4, m5, m6, m7, m8,
+		m9, m10, m11, m12, m13, m14, m15;
+
+	/* (PMU) Load accumulated path metrics */
+	m0 = _mm_load_si128((__m128i *) &sums[0]);
+	m1 = _mm_load_si128((__m128i *) &sums[8]);
+	m2 = _mm_load_si128((__m128i *) &sums[16]);
+	m3 = _mm_load_si128((__m128i *) &sums[24]);
+	m4 = _mm_load_si128((__m128i *) &sums[32]);
+	m5 = _mm_load_si128((__m128i *) &sums[40]);
+	m6 = _mm_load_si128((__m128i *) &sums[48]);
+	m7 = _mm_load_si128((__m128i *) &sums[56]);
+
+	/* (PMU) Deinterleave to even-odd registers */
+	SSE_DEINTERLEAVE_K7(m0, m1, m2, m3 ,m4 ,m5, m6, m7,
+			    m8, m9, m10, m11, m12, m13, m14, m15)
+
+	/* (BMU) Load input symbols */
+	m7 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
+
+	/* (BMU) Load trellis outputs */
+	m0 = _mm_load_si128((__m128i *) &out[0]);
+	m1 = _mm_load_si128((__m128i *) &out[8]);
+	m2 = _mm_load_si128((__m128i *) &out[16]);
+	m3 = _mm_load_si128((__m128i *) &out[24]);
+
+	SSE_BRANCH_METRIC_N2(m0, m1, m2, m3, m7, m4, m5)
+
+	m0 = _mm_load_si128((__m128i *) &out[32]);
+	m1 = _mm_load_si128((__m128i *) &out[40]);
+	m2 = _mm_load_si128((__m128i *) &out[48]);
+	m3 = _mm_load_si128((__m128i *) &out[56]);
+
+	SSE_BRANCH_METRIC_N2(m0, m1, m2, m3, m7, m6, m7)
+
+	/* (PMU) Butterflies: 0-15 */
+	SSE_BUTTERFLY(m8, m9, m4, m0, m1)
+	SSE_BUTTERFLY(m10, m11, m5, m2, m3)
+
+	_mm_store_si128((__m128i *) &paths[0], m0);
+	_mm_store_si128((__m128i *) &paths[8], m2);
+	_mm_store_si128((__m128i *) &paths[32], m9);
+	_mm_store_si128((__m128i *) &paths[40], m11);
+
+	/* (PMU) Butterflies: 17-31 */
+	SSE_BUTTERFLY(m12, m13, m6, m0, m2)
+	SSE_BUTTERFLY(m14, m15, m7, m9, m11)
+
+	_mm_store_si128((__m128i *) &paths[16], m0);
+	_mm_store_si128((__m128i *) &paths[24], m9);
+	_mm_store_si128((__m128i *) &paths[48], m13);
+	_mm_store_si128((__m128i *) &paths[56], m15);
+
+	if (norm)
+		SSE_NORMALIZE_K7(m4, m1, m5, m3, m6, m2,
+				 m7, m11, m0, m8, m9, m10)
+
+	_mm_store_si128((__m128i *) &sums[0], m4);
+	_mm_store_si128((__m128i *) &sums[8], m5);
+	_mm_store_si128((__m128i *) &sums[16], m6);
+	_mm_store_si128((__m128i *) &sums[24], m7);
+	_mm_store_si128((__m128i *) &sums[32], m1);
+	_mm_store_si128((__m128i *) &sums[40], m3);
+	_mm_store_si128((__m128i *) &sums[48], m2);
+	_mm_store_si128((__m128i *) &sums[56], m11);
+}
+
+/* Combined BMU/PMU (K=7, N=3 and N=4)
+ * Compute branch metrics followed by path metrics for half rate 64-state
+ * trellis. 32 butterfly operations are computed. Deinterleave path
+ * metrics before computing branch metrics as in the half rate case.
+ */
+__always_inline static void _sse_metrics_k7_n4(const int16_t *val,
+	const int16_t *out, int16_t *sums, int16_t *paths, int norm)
+{
+	__m128i m0, m1, m2, m3, m4, m5, m6, m7;
+	__m128i m8, m9, m10, m11, m12, m13, m14, m15;
+
+	/* (PMU) Load accumulated path metrics */
+	m0 = _mm_load_si128((__m128i *) &sums[0]);
+	m1 = _mm_load_si128((__m128i *) &sums[8]);
+	m2 = _mm_load_si128((__m128i *) &sums[16]);
+	m3 = _mm_load_si128((__m128i *) &sums[24]);
+	m4 = _mm_load_si128((__m128i *) &sums[32]);
+	m5 = _mm_load_si128((__m128i *) &sums[40]);
+	m6 = _mm_load_si128((__m128i *) &sums[48]);
+	m7 = _mm_load_si128((__m128i *) &sums[56]);
+
+	/* (PMU) Deinterleave into even and odd packed registers */
+	SSE_DEINTERLEAVE_K7(m0, m1, m2, m3 ,m4 ,m5, m6, m7,
+			    m8, m9, m10, m11, m12, m13, m14, m15)
+
+	/* (BMU) Load and expand 8-bit input out to 16-bits */
+	m7 = _mm_castpd_si128(_mm_loaddup_pd((double const *) val));
+
+	/* (BMU) Load and compute branch metrics */
+	m0 = _mm_load_si128((__m128i *) &out[0]);
+	m1 = _mm_load_si128((__m128i *) &out[8]);
+	m2 = _mm_load_si128((__m128i *) &out[16]);
+	m3 = _mm_load_si128((__m128i *) &out[24]);
+
+	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m4)
+
+	m0 = _mm_load_si128((__m128i *) &out[32]);
+	m1 = _mm_load_si128((__m128i *) &out[40]);
+	m2 = _mm_load_si128((__m128i *) &out[48]);
+	m3 = _mm_load_si128((__m128i *) &out[56]);
+
+	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m5)
+
+	m0 = _mm_load_si128((__m128i *) &out[64]);
+	m1 = _mm_load_si128((__m128i *) &out[72]);
+	m2 = _mm_load_si128((__m128i *) &out[80]);
+	m3 = _mm_load_si128((__m128i *) &out[88]);
+
+	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m6)
+
+	m0 = _mm_load_si128((__m128i *) &out[96]);
+	m1 = _mm_load_si128((__m128i *) &out[104]);
+	m2 = _mm_load_si128((__m128i *) &out[112]);
+	m3 = _mm_load_si128((__m128i *) &out[120]);
+
+	SSE_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m7)
+
+	/* (PMU) Butterflies: 0-15 */
+	SSE_BUTTERFLY(m8, m9, m4, m0, m1)
+	SSE_BUTTERFLY(m10, m11, m5, m2, m3)
+
+	_mm_store_si128((__m128i *) &paths[0], m0);
+	_mm_store_si128((__m128i *) &paths[8], m2);
+	_mm_store_si128((__m128i *) &paths[32], m9);
+	_mm_store_si128((__m128i *) &paths[40], m11);
+
+	/* (PMU) Butterflies: 17-31 */
+	SSE_BUTTERFLY(m12, m13, m6, m0, m2)
+	SSE_BUTTERFLY(m14, m15, m7, m9, m11)
+
+	_mm_store_si128((__m128i *) &paths[16], m0);
+	_mm_store_si128((__m128i *) &paths[24], m9);
+	_mm_store_si128((__m128i *) &paths[48], m13);
+	_mm_store_si128((__m128i *) &paths[56], m15);
+
+	if (norm)
+		SSE_NORMALIZE_K7(m4, m1, m5, m3, m6, m2,
+				 m7, m11, m0, m8, m9, m10)
+
+	_mm_store_si128((__m128i *) &sums[0], m4);
+	_mm_store_si128((__m128i *) &sums[8], m5);
+	_mm_store_si128((__m128i *) &sums[16], m6);
+	_mm_store_si128((__m128i *) &sums[24], m7);
+	_mm_store_si128((__m128i *) &sums[32], m1);
+	_mm_store_si128((__m128i *) &sums[40], m3);
+	_mm_store_si128((__m128i *) &sums[48], m2);
+	_mm_store_si128((__m128i *) &sums[56], m11);
+}