Forwards and backwards working

10 files changed, 134 insertions, 93 deletions

diff --git a/Makefile.am b/Makefile.am
index 27f515f..33b171c 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -1,3 +1,3 @@
 AUTOMAKE_OPTIONS = foreign
-SUBDIRS = src 
+SUBDIRS = src tests 
 
diff --git a/Makefile.in b/Makefile.in
index b63edd2..a889030 100644
--- a/Makefile.in
+++ b/Makefile.in
@@ -217,7 +217,7 @@ top_build_prefix = @top_build_prefix@
 top_builddir = @top_builddir@
 top_srcdir = @top_srcdir@
 AUTOMAKE_OPTIONS = foreign
-SUBDIRS = src 
+SUBDIRS = src tests 
 all: config.h
 	$(MAKE) $(AM_MAKEFLAGS) all-recursive
 
diff --git a/configure b/configure
index 81fd0eb..d412018 100755
--- a/configure
+++ b/configure
@@ -5049,7 +5049,7 @@ done
 
 
 
-ac_config_files="$ac_config_files Makefile src/Makefile"
+ac_config_files="$ac_config_files Makefile src/Makefile tests/Makefile"
 
 cat >confcache <<\_ACEOF
 # This file is a shell script that caches the results of configure
@@ -5784,6 +5784,7 @@ do
     "depfiles") CONFIG_COMMANDS="$CONFIG_COMMANDS depfiles" ;;
     "Makefile") CONFIG_FILES="$CONFIG_FILES Makefile" ;;
     "src/Makefile") CONFIG_FILES="$CONFIG_FILES src/Makefile" ;;
+    "tests/Makefile") CONFIG_FILES="$CONFIG_FILES tests/Makefile" ;;
 
   *) as_fn_error $? "invalid argument: \`$ac_config_target'" "$LINENO" 5;;
   esac
diff --git a/configure.ac b/configure.ac
index 775ae7b..06cb913 100644
--- a/configure.ac
+++ b/configure.ac
@@ -55,5 +55,6 @@ AC_CHECK_FUNCS([gettimeofday pow])
 
 AC_CONFIG_FILES([Makefile
                  src/Makefile
+                 tests/Makefile
                  ])
 AC_OUTPUT
diff --git a/include/ffts.h b/include/ffts.h
index 1431cf6..e266491 100644
--- a/include/ffts.h
+++ b/include/ffts.h
@@ -37,11 +37,26 @@
 #include <complex.h>
 #include <math.h>
 #include <stdint.h>
+#include <stddef.h>
 
-typedef struct ffts_plan_t;
+typedef size_t transform_index_t;
+
+struct _ffts_plan_t {
+	ptrdiff_t *is;
+	ptrdiff_t *offsets;
+	void __attribute__ ((aligned(32))) **ws;
+	void (*firstpass)(const float * restrict, float * restrict, size_t, struct _ffts_plan_t * restrict);
+	size_t i0, i1, i2;
+	uint64_t n_bits, leaftime;
+	
+	transform_index_t *transforms; 
+};
+
+
+typedef struct _ffts_plan_t ffts_plan_t;
 
 void ffts_execute(ffts_plan_t * restrict, const void * restrict, const void * restrict);
-ffts_plan_t *ffts_init(size_t N);
+ffts_plan_t *ffts_init(size_t N, int sign);
 
 
 #endif
diff --git a/src/Makefile.in b/src/Makefile.in
index 73fecf8..fbc26bb 100644
--- a/src/Makefile.in
+++ b/src/Makefile.in
@@ -171,9 +171,9 @@ $(srcdir)/Makefile.in:  $(srcdir)/Makefile.am  $(am__configure_deps)
 	      exit 1;; \
 	  esac; \
 	done; \
-	echo ' cd $(top_srcdir) && $(AUTOMAKE) --foreign src/Makefile'; \
+	echo ' cd $(top_srcdir) && $(AUTOMAKE) --gnu src/Makefile'; \
 	$(am__cd) $(top_srcdir) && \
-	  $(AUTOMAKE) --foreign src/Makefile
+	  $(AUTOMAKE) --gnu src/Makefile
 .PRECIOUS: Makefile
 Makefile: $(srcdir)/Makefile.in $(top_builddir)/config.status
 	@case '$?' in \
diff --git a/src/cp_sse.c b/src/cp_sse.c
index f59c60b..8b09031 100644
--- a/src/cp_sse.c
+++ b/src/cp_sse.c
@@ -35,23 +35,26 @@ firstpass_64(const float * restrict in, float * restrict out, size_t N, ffts_pla
 void 
 firstpass_32(const data_t * restrict in, data_t * restrict out, size_t N, ffts_plan_t * restrict p) {
   __m128 r0_1,r2_3,r4_5,r6_7,r8_9,r10_11,r12_13,r14_15,r16_17,r18_19,r20_21,r22_23,r24_25,r26_27,r28_29,r30_31;
+  float *LUT8 = p->ws[0];
+  float *LUT16 = p->ws[1];
+  float *LUT32 = p->ws[2];
 
   L_4_4(in+0,in+32,in+16,in+48,&r0_1,&r2_3,&r16_17,&r18_19);
   L_2_2(in+8,in+40,in+56,in+24,&r4_5,&r6_7,&r20_21,&r22_23);
-  K_N(VLIT4(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),VLIT4(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,-0,0),&r0_1,&r2_3,&r4_5,&r6_7);
+  K_N(_mm_load_ps(LUT8),_mm_load_ps(LUT8+4),&r0_1,&r2_3,&r4_5,&r6_7);
   L_4_2(in+4,in+36,in+20,in+52,&r8_9,&r10_11,&r28_29,&r30_31);
   L_4_4(in+60,in+28,in+12,in+44,&r12_13,&r14_15,&r24_25,&r26_27);
-  K_N(VLIT4(0.92387953251128673848313610506011,0.92387953251128673848313610506011,1,1),VLIT4(0.38268343236508978177923268049199,-0.38268343236508978177923268049199,-0,0),&r0_1,&r4_5,&r8_9,&r12_13);
-  K_N(VLIT4(0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.70710678118654757273731092936941,0.70710678118654757273731092936941),VLIT4(0.92387953251128673848313610506011,-0.92387953251128673848313610506011,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r2_3,&r6_7,&r10_11,&r14_15);
-  K_N(VLIT4(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),VLIT4(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,-0,0),&r16_17,&r18_19,&r20_21,&r22_23);
-  K_N(VLIT4(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),VLIT4(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,-0,0),&r24_25,&r26_27,&r28_29,&r30_31);
-  K_N(VLIT4(0.98078528040323043057924223830923,0.98078528040323043057924223830923,1,1),VLIT4(0.19509032201612824808378832130984,-0.19509032201612824808378832130984,-0,0),&r0_1,&r8_9,&r16_17,&r24_25);
+  K_N(_mm_load_ps(LUT16),_mm_load_ps(LUT16+4),&r0_1,&r4_5,&r8_9,&r12_13);
+  K_N(_mm_load_ps(LUT16+8),_mm_load_ps(LUT16+12),&r2_3,&r6_7,&r10_11,&r14_15);
+  K_N(_mm_load_ps(LUT8),_mm_load_ps(LUT8+4),&r16_17,&r18_19,&r20_21,&r22_23);
+  K_N(_mm_load_ps(LUT8),_mm_load_ps(LUT8+4),&r24_25,&r26_27,&r28_29,&r30_31);
+  K_N(_mm_load_ps(LUT32),_mm_load_ps(LUT32+4),&r0_1,&r8_9,&r16_17,&r24_25);
   S_4(r0_1,r8_9,r16_17,r24_25,out+0,out+16,out+32,out+48);
-  K_N(VLIT4(0.8314696123025452356714026791451,0.8314696123025452356714026791451,0.92387953251128673848313610506011,0.92387953251128673848313610506011),VLIT4(0.55557023301960217764872140833177,-0.55557023301960217764872140833177,0.38268343236508978177923268049199,-0.38268343236508978177923268049199),&r2_3,&r10_11,&r18_19,&r26_27);
+  K_N(_mm_load_ps(LUT32+8),_mm_load_ps(LUT32+12),&r2_3,&r10_11,&r18_19,&r26_27);
   S_4(r2_3,r10_11,r18_19,r26_27,out+4,out+20,out+36,out+52);
-  K_N(VLIT4(0.55557023301960228867102387084742,0.55557023301960228867102387084742,0.70710678118654757273731092936941,0.70710678118654757273731092936941),VLIT4(0.83146961230254512464910021662945,-0.83146961230254512464910021662945,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r4_5,&r12_13,&r20_21,&r28_29);
+  K_N(_mm_load_ps(LUT32+16),_mm_load_ps(LUT32+20),&r4_5,&r12_13,&r20_21,&r28_29);
   S_4(r4_5,r12_13,r20_21,r28_29,out+8,out+24,out+40,out+56);
-  K_N(VLIT4(0.19509032201612830359493955256767,0.19509032201612830359493955256767,0.38268343236508983729038391174981,0.38268343236508983729038391174981),VLIT4(0.98078528040323043057924223830923,-0.98078528040323043057924223830923,0.92387953251128673848313610506011,-0.92387953251128673848313610506011),&r6_7,&r14_15,&r22_23,&r30_31);
+  K_N(_mm_load_ps(LUT32+24),_mm_load_ps(LUT32+28),&r6_7,&r14_15,&r22_23,&r30_31);
   S_4(r6_7,r14_15,r22_23,r30_31,out+12,out+28,out+44,out+60);
 
 }
@@ -59,21 +62,24 @@ firstpass_32(const data_t * restrict in, data_t * restrict out, size_t N, ffts_p
 void 
 firstpass_16(const data_t * restrict in, data_t * restrict out, size_t N, ffts_plan_t * restrict p) {
   __m128 r0_1,r2_3,r4_5,r6_7,r8_9,r10_11,r12_13,r14_15;
+  float *LUT8 = p->ws[0];
+  float *LUT16 = p->ws[1];
 
   L_4_4(in+0,in+16,in+8,in+24,&r0_1,&r2_3,&r8_9,&r10_11);
   L_2_4(in+4,in+20,in+28,in+12,&r4_5,&r6_7,&r14_15,&r12_13);
-  K_N(VLIT4(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),VLIT4(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,-0,0),&r0_1,&r2_3,&r4_5,&r6_7);
-  K_N(VLIT4(0.92387953251128673848313610506011,0.92387953251128673848313610506011,1,1),VLIT4(0.38268343236508978177923268049199,-0.38268343236508978177923268049199,-0,0),&r0_1,&r4_5,&r8_9,&r12_13);
+  K_N(_mm_load_ps(LUT8),_mm_load_ps(LUT8+4),&r0_1,&r2_3,&r4_5,&r6_7);
+  K_N(_mm_load_ps(LUT16),_mm_load_ps(LUT16+4),&r0_1,&r4_5,&r8_9,&r12_13);
   S_4(r0_1,r4_5,r8_9,r12_13,out+0,out+8,out+16,out+24);
-  K_N(VLIT4(0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.70710678118654757273731092936941,0.70710678118654757273731092936941),VLIT4(0.92387953251128673848313610506011,-0.92387953251128673848313610506011,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r2_3,&r6_7,&r10_11,&r14_15);
+  K_N(_mm_load_ps(LUT16+8),_mm_load_ps(LUT16+12),&r2_3,&r6_7,&r10_11,&r14_15);
   S_4(r2_3,r6_7,r10_11,r14_15,out+4,out+12,out+20,out+28);
-
 }
+
 void 
 firstpass_8(const data_t * restrict in, data_t * restrict out, size_t N, ffts_plan_t * restrict p) {
   __m128 r0_1,r2_3,r4_5,r6_7;
-  L_4_2(in+0,in+8,in+4,in+12,&r0_1,&r2_3,&r4_5,&r6_7);
-  K_N(VLIT4(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),VLIT4(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,-0,0),&r0_1,&r2_3,&r4_5,&r6_7);
+  float *LUT8 = p->ws[0];
+	L_4_2(in+0,in+8,in+4,in+12,&r0_1,&r2_3,&r4_5,&r6_7);
+  K_N(_mm_load_ps(LUT8),_mm_load_ps(LUT8+4),&r0_1,&r2_3,&r4_5,&r6_7);
   S_4(r0_1,r2_3,r4_5,r6_7,out+0,out+4,out+8,out+12);
 }
 void 
@@ -165,10 +171,13 @@ void ffts_execute(ffts_plan_t *p, const void * restrict in, void * restrict out,
 }
 
 
-ffts_plan_t *ffts_init(size_t N) {
+ffts_plan_t *ffts_init(size_t N, int sign) {
 	ffts_plan_t *p = malloc(sizeof(ffts_plan_t));
 	size_t leafN = 16;	
 	size_t i;	
+		
+	if(sign < 0) MULI_SIGN = _mm_set_ps(-0.0f, 0.0f, -0.0f, 0.0f);
+	else         MULI_SIGN = _mm_set_ps(0.0f, -0.0f, 0.0f, -0.0f);
 	
 	if(N > 32) {
 		init_offsets(p, N, leafN);
@@ -178,25 +187,68 @@ ffts_plan_t *ffts_init(size_t N) {
 		if(N == 64) p->firstpass = &firstpass_64;
 		else if(__builtin_ctzl(N) & 1) p->firstpass = &firstpass_type_2;	
 		else p->firstpass = &firstpass_type_1;	
-		/*      LUTS           */
 
-		size_t n_luts = __builtin_ctzl(N/leafN);
+		LEAFLUT[0] = _mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941);
+		LEAFLUT[1] = _mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0.70710678118654746171500846685376,-0.70710678118654746171500846685376);
+		LEAFLUT[2] = _mm_set_ps(0.92387953251128673848313610506011,0.92387953251128673848313610506011,0.92387953251128673848313610506011,0.92387953251128673848313610506011);
+		LEAFLUT[3] = _mm_set_ps(0.38268343236508978177923268049199,-0.38268343236508978177923268049199,0.38268343236508978177923268049199,-0.38268343236508978177923268049199);
+		LEAFLUT[4] = _mm_set_ps(0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.38268343236508983729038391174981);
+		LEAFLUT[5] = _mm_set_ps(0.92387953251128673848313610506011,-0.92387953251128673848313610506011,0.92387953251128673848313610506011,-0.92387953251128673848313610506011);
+  	
+		LEAFLUT[6] = _mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1);
+		LEAFLUT[7] = _mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0,-0);
+		LEAFLUT[8] = _mm_set_ps(0.92387953251128673848313610506011,0.92387953251128673848313610506011,1,1);
+		LEAFLUT[9] = _mm_set_ps(0.38268343236508978177923268049199,-0.38268343236508978177923268049199,0,-0);
+		LEAFLUT[10] = _mm_set_ps(0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.70710678118654757273731092936941,0.70710678118654757273731092936941);
+		LEAFLUT[11] = _mm_set_ps(0.92387953251128673848313610506011,-0.92387953251128673848313610506011,0.70710678118654746171500846685376,-0.70710678118654746171500846685376);
 
+		
+		if(sign > 0) {
+			LEAFLUT[1] = _mm_xor_ps(LEAFLUT[1], _mm_set_ps(-0.0f,-0.0f,-0.0f,-0.0f));
+			LEAFLUT[3] = _mm_xor_ps(LEAFLUT[3], _mm_set_ps(-0.0f,-0.0f,-0.0f,-0.0f));
+			LEAFLUT[5] = _mm_xor_ps(LEAFLUT[5], _mm_set_ps(-0.0f,-0.0f,-0.0f,-0.0f));
+			LEAFLUT[7] = _mm_xor_ps(LEAFLUT[7], _mm_set_ps(-0.0f,-0.0f,-0.0f,-0.0f));
+			LEAFLUT[9] = _mm_xor_ps(LEAFLUT[9], _mm_set_ps(-0.0f,-0.0f,-0.0f,-0.0f));
+			LEAFLUT[11] = _mm_xor_ps(LEAFLUT[11], _mm_set_ps(-0.0f,-0.0f,-0.0f,-0.0f));
+		}
+		
 		p->i0 = N/leafN/3+1;
 		p->i1 = N/leafN/3;
 		if((N/leafN) % 3 > 1) p->i1++;
 		p->i0/=2;
 		p->i1/=2;
 
+	}else{
+		p->transforms = malloc(2 * sizeof(transform_index_t));
+		p->transforms[0] = 0;
+		p->transforms[1] = 1;
+		if(N == 2) p->firstpass = &firstpass_2;
+		else if(N == 4) p->firstpass = &firstpass_4;
+		else if(N == 8) p->firstpass = &firstpass_8;
+		else if(N == 16) p->firstpass = &firstpass_16;
+		else if(N == 32) p->firstpass = &firstpass_32;
+
+	}
+
+		int hardcoded = 0;
 
-		//	printf("n_luts = %zu\n", n_luts);
+		/*      LUTS           */
+		size_t n_luts = __builtin_ctzl(N/leafN);
+		if(N <= 32) { n_luts = __builtin_ctzl(N/4); hardcoded = 1; }
+
+
+		//printf("n_luts = %zu\n", n_luts);
 		p->ws = malloc(n_luts * sizeof(data_t *));
 		cdata_t *w;
 
 		int n = leafN*2;
+		if(hardcoded) n = 8;
+
 		for(i=0;i<n_luts;i++) {
-			//	printf("LUT[%zu] = %d\n", i, n);	
-			if(!i) {
+				
+			//printf("LUT[%zu] = %d\n", i, n);	
+			
+			if(!i || hardcoded) {
 
 				w = _mm_malloc(n/4 * 2 * sizeof(cdata_t), 32);
 
@@ -216,15 +268,15 @@ ffts_plan_t *ffts_init(size_t N) {
 					__m128 re, im;
 					re = _mm_shuffle_ps(temp0, temp0, _MM_SHUFFLE(2, 2, 0, 0));
 					im = _mm_shuffle_ps(temp0, temp0, _MM_SHUFFLE(3, 3, 1, 1));
-					im = _mm_xor_ps(im, _mm_set_ps(-0.0f, 0.0f, -0.0f, 0.0f));
+					im = _mm_xor_ps(im, MULI_SIGN);
 					_mm_store_ps(fw + j*4  , re);
 					_mm_store_ps(fw + j*4+4, im);
 				}
 
-				//	for(j=0;j<n/2;j++) {
-				//		printf("%f %f\n", creal(w[j]), cimag(w[j]));
+	//		  	for(j=0;j<n/2;j++) {
+	//		  		printf("%f %f\n", creal(w[j]), cimag(w[j]));
 
-				//	}
+	//		  	}
 
 				_mm_free(w0);
 			}else{
@@ -252,21 +304,21 @@ ffts_plan_t *ffts_init(size_t N) {
 					temp0 = _mm_load_ps(fw0 + j*2);
 					re = _mm_shuffle_ps(temp0, temp0, _MM_SHUFFLE(2, 2, 0, 0));
 					im = _mm_shuffle_ps(temp0, temp0, _MM_SHUFFLE(3, 3, 1, 1));
-					im = _mm_xor_ps(im, _mm_set_ps(-0.0f, 0.0f, -0.0f, 0.0f));
+					im = _mm_xor_ps(im, MULI_SIGN);
 					_mm_store_ps(fw + j*2*6  , re);
 					_mm_store_ps(fw + j*2*6+4, im);
 
 					temp1 = _mm_load_ps(fw1 + j*2);
 					re = _mm_shuffle_ps(temp1, temp1, _MM_SHUFFLE(2, 2, 0, 0));
 					im = _mm_shuffle_ps(temp1, temp1, _MM_SHUFFLE(3, 3, 1, 1));
-					im = _mm_xor_ps(im, _mm_set_ps(-0.0f, 0.0f, -0.0f, 0.0f));
+					im = _mm_xor_ps(im, MULI_SIGN);
 					_mm_store_ps(fw + j*2*6+8 , re);
 					_mm_store_ps(fw + j*2*6+12, im);
 
 					temp2 = _mm_load_ps(fw2 + j*2);
 					re = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(2, 2, 0, 0));
 					im = _mm_shuffle_ps(temp2, temp2, _MM_SHUFFLE(3, 3, 1, 1));
-					im = _mm_xor_ps(im, _mm_set_ps(-0.0f, 0.0f, -0.0f, 0.0f));
+					im = _mm_xor_ps(im, MULI_SIGN);
 					_mm_store_ps(fw + j*2*6+16, re);
 					_mm_store_ps(fw + j*2*6+20, im);
 				}
@@ -280,22 +332,11 @@ ffts_plan_t *ffts_init(size_t N) {
 			n *= 2;
 		}
 
-		p->n_bits = log(N)/log(2) - log(leafN*2)/log(2);
-	}else{
-		p->transforms = malloc(2 * sizeof(transform_index_t));
-		p->transforms[0] = 0;
-		p->transforms[1] = 1;
-		if(N == 2) p->firstpass = &firstpass_2;
-		else if(N == 4) p->firstpass = &firstpass_4;
-		else if(N == 8) p->firstpass = &firstpass_8;
-		else if(N == 16) p->firstpass = &firstpass_16;
-		else if(N == 32) p->firstpass = &firstpass_32;
-
-	}
 
+	
 	return p;
 }
-
+/*
 int main(int argc, char *argv[]) {
 	int n = atoi(argv[1]);
 	int count = atoi(argv[2]);
@@ -341,4 +382,4 @@ int main(int argc, char *argv[]) {
 	printf("Time: %f seconds, CTGs: %f Leaftime: %f \n", tt, ctgs, lt);
 
 	return 0;
-}
+}*/
diff --git a/src/cp_sse.h b/src/cp_sse.h
index c6d15dc..2c6825f 100644
--- a/src/cp_sse.h
+++ b/src/cp_sse.h
@@ -8,7 +8,6 @@
 #include <stddef.h>
 #include <xmmintrin.h>
 #include <stdint.h>
-#include <mach/mach_time.h>
 
 typedef complex float cdata_t;
 typedef float data_t;
@@ -22,19 +21,11 @@ struct _ffts_plan_t {
 	ptrdiff_t *offsets;
 	void __attribute__ ((aligned(32))) **ws;
 	void (*firstpass)(const float * restrict, float * restrict, size_t, struct _ffts_plan_t * restrict);
-	size_t i0, i1, i2;
-	uint64_t n_bits, leaftime;
+	size_t i0, i1;
 	
 	transform_index_t *transforms; 
 };
 
-
 typedef struct _ffts_plan_t ffts_plan_t;
 
-
-typedef struct _split_vec_t {
-	__m128 re, im;
-} split_vec_t;
-
-
 #endif
diff --git a/src/macros.h b/src/macros.h
index 0a84802..b2f44e6 100644
--- a/src/macros.h
+++ b/src/macros.h
@@ -7,8 +7,10 @@
 
 #define VLIT4 _mm_set_ps
 
+__m128 MULI_SIGN;
+
 __INLINE __m128 IMULI(__m128 a) {
-	__m128 temp = _mm_xor_ps(a, _mm_set_ps(-0.0f, 0.0f, -0.0f, 0.0f));
+	__m128 temp = _mm_xor_ps(a, MULI_SIGN);//_mm_set_ps(-0.0f, 0.0f, -0.0f, 0.0f));
 	return _mm_shuffle_ps(temp, temp, _MM_SHUFFLE(2,3,0,1));
 }
 
@@ -113,6 +115,8 @@ __INLINE void TX2(__m128 *a, __m128 *b) {
     *a = TX2_t0; *b = TX2_t1; 
 }
 
+__m128 __attribute__((aligned(32))) LEAFLUT[12];
+
 __INLINE void 
 LEAF_EE(size_t ** restrict is, const data_t * restrict in, size_t ** restrict out_offsets, data_t * restrict out) {
   __m128 r0,r1,r2,r3,r4,r5,r6,r7,r8,r9,r10,r11,r12,r13,r14,r15;
@@ -124,23 +128,17 @@ LEAF_EE(size_t ** restrict is, const data_t * restrict in, size_t ** restrict ou
   L_4(in+(*is)[0],in+(*is)[1],in+(*is)[2],in+(*is)[3],&r0,&r1,&r2,&r3);
   L_2(in+(*is)[4],in+(*is)[5],in+(*is)[6],in+(*is)[7],&r4,&r5,&r6,&r7);
   K_0(&r0,&r2,&r4,&r6);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r1,&r3,&r5,&r7);
+  K_N(LEAFLUT[0],LEAFLUT[1],&r1,&r3,&r5,&r7);
   L_4(in+(*is)[8],in+(*is)[9],in+(*is)[10],in+(*is)[11],&r8,&r9,&r10,&r11);
   L_4(in+(*is)[12],in+(*is)[13],in+(*is)[14],in+(*is)[15],&r12,&r13,&r14,&r15);
   K_0(&r0,&r4,&r8,&r12);
-  K_N(_mm_set_ps(0.92387953251128673848313610506011,0.92387953251128673848313610506011,0.92387953251128673848313610506011,0.92387953251128673848313610506011),_mm_set_ps(0.38268343236508978177923268049199,-0.38268343236508978177923268049199,0.38268343236508978177923268049199,-0.38268343236508978177923268049199),&r1,&r5,&r9,&r13);
-  TX2(&r0,&r1);
-  TX2(&r4,&r5);
-  TX2(&r8,&r9);
-  TX2(&r12,&r13);
+  K_N(LEAFLUT[2],LEAFLUT[3],&r1,&r5,&r9,&r13);
+  TX2(&r0,&r1); TX2(&r4,&r5); TX2(&r8,&r9); TX2(&r12,&r13);
   S_4(r0,r4,r8,r12,out0+0,out0+8,out0+16,out0+24);
   S_4(r1,r5,r9,r13,out1+0,out1+8,out1+16,out1+24);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r2,&r6,&r10,&r14);
-  K_N(_mm_set_ps(0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.38268343236508983729038391174981),_mm_set_ps(0.92387953251128673848313610506011,-0.92387953251128673848313610506011,0.92387953251128673848313610506011,-0.92387953251128673848313610506011),&r3,&r7,&r11,&r15);
-  TX2(&r2,&r3);
-  TX2(&r6,&r7);
-  TX2(&r10,&r11);
-  TX2(&r14,&r15);
+  K_N(LEAFLUT[0],LEAFLUT[1],&r2,&r6,&r10,&r14);
+  K_N(LEAFLUT[4],LEAFLUT[5],&r3,&r7,&r11,&r15);
+  TX2(&r2,&r3); TX2(&r6,&r7); TX2(&r10,&r11); TX2(&r14,&r15);
   S_4(r2,r6,r10,r14,out0+4,out0+12,out0+20,out0+28);
   S_4(r3,r7,r11,r15,out1+4,out1+12,out1+20,out1+28);
 	*is += 16;
@@ -158,21 +156,15 @@ LEAF_OO(size_t ** restrict is, const data_t * restrict in, size_t ** restrict ou
   L_4(in+(*is)[0],in+(*is)[1],in+(*is)[2],in+(*is)[3],&r0,&r1,&r2,&r3);
   L_2(in+(*is)[4],in+(*is)[5],in+(*is)[6],in+(*is)[7],&r4,&r5,&r6,&r7);
   K_0(&r0,&r2,&r4,&r6);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r1,&r3,&r5,&r7);
-  TX2(&r0,&r1);
-  TX2(&r2,&r3);
-  TX2(&r4,&r5);
-  TX2(&r6,&r7);
+  K_N(LEAFLUT[0],LEAFLUT[1],&r1,&r3,&r5,&r7);
+  TX2(&r0,&r1); TX2(&r2,&r3); TX2(&r4,&r5); TX2(&r6,&r7);
   S_4(r0,r2,r4,r6,out0+0,out0+4,out0+8,out0+12);
   S_4(r1,r3,r5,r7,out1+0,out1+4,out1+8,out1+12);
   L_4(in+(*is)[8],in+(*is)[9],in+(*is)[10],in+(*is)[11],&r8,&r9,&r10,&r11);
   L_2(in+(*is)[12],in+(*is)[13],in+(*is)[14],in+(*is)[15],&r12,&r13,&r14,&r15);
   K_0(&r8,&r10,&r12,&r14);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941,0.70710678118654757273731092936941),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r9,&r11,&r13,&r15);
-  TX2(&r8,&r9);
-  TX2(&r10,&r11);
-  TX2(&r12,&r13);
-  TX2(&r14,&r15);
+  K_N(LEAFLUT[0],LEAFLUT[1],&r9,&r11,&r13,&r15);
+  TX2(&r8,&r9); TX2(&r10,&r11); TX2(&r12,&r13); TX2(&r14,&r15);
   S_4(r8,r10,r12,r14,out0+16,out0+20,out0+24,out0+28);
   S_4(r9,r11,r13,r15,out1+16,out1+20,out1+24,out1+28);
 
@@ -264,16 +256,16 @@ LEAF_OE(size_t ** restrict is, const data_t * restrict in, size_t ** restrict ou
   
 	L_4_4(in+(*is)[0],in+(*is)[1],in+(*is)[2],in+(*is)[3],&r0_1,&r2_3,&r24_25,&r26_27);
   L_2_4(in+(*is)[4],in+(*is)[5],in+(*is)[6],in+(*is)[7],&r4_5,&r6_7,&r30_31,&r28_29);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0,-0),&r0_1,&r2_3,&r4_5,&r6_7);
+  K_N(LEAFLUT[6],LEAFLUT[7],&r0_1,&r2_3,&r4_5,&r6_7);
   S_4(r0_1,r2_3,r4_5,r6_7,out0+0,out0+4,out0+8,out0+12);
   L_4_4(in+(*is)[8],in+(*is)[9],in+(*is)[10],in+(*is)[11],&r8_9,&r10_11,&r16_17,&r18_19);
   L_2_2(in+(*is)[12],in+(*is)[13],in+(*is)[14],in+(*is)[15],&r12_13,&r14_15,&r20_21,&r22_23);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0,-0),&r8_9,&r10_11,&r12_13,&r14_15);
+  K_N(LEAFLUT[6],LEAFLUT[7],&r8_9,&r10_11,&r12_13,&r14_15);
   S_4(r8_9,r10_11,r12_13,r14_15,out0+16,out0+20,out0+24,out0+28);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0,-0),&r16_17,&r18_19,&r20_21,&r22_23);
-  K_N(_mm_set_ps(0.92387953251128673848313610506011,0.92387953251128673848313610506011,1,1),_mm_set_ps(0.38268343236508978177923268049199,-0.38268343236508978177923268049199,0,-0),&r16_17,&r20_21,&r24_25,&r28_29);
+  K_N(LEAFLUT[6],LEAFLUT[7],&r16_17,&r18_19,&r20_21,&r22_23);
+  K_N(LEAFLUT[8],LEAFLUT[9],&r16_17,&r20_21,&r24_25,&r28_29);
   S_4(r16_17,r20_21,r24_25,r28_29,out1+0,out1+8,out1+16,out1+24);
-  K_N(_mm_set_ps(0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.70710678118654757273731092936941,0.70710678118654757273731092936941),_mm_set_ps(0.92387953251128673848313610506011,-0.92387953251128673848313610506011,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r18_19,&r22_23,&r26_27,&r30_31);
+  K_N(LEAFLUT[10],LEAFLUT[11],&r18_19,&r22_23,&r26_27,&r30_31);
   S_4(r18_19,r22_23,r26_27,r30_31,out1+4,out1+12,out1+20,out1+28);
 
 	*is += 16;
@@ -289,16 +281,16 @@ LEAF_EO(size_t ** restrict is, const data_t * restrict in, size_t ** restrict ou
 
   L_4_4(in+(*is)[0],in+(*is)[1],in+(*is)[2],in+(*is)[3],&r0_1,&r2_3,&r16_17,&r18_19);
   L_2_2(in+(*is)[4],in+(*is)[5],in+(*is)[6],in+(*is)[7],&r4_5,&r6_7,&r20_21,&r22_23);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0,-0),&r0_1,&r2_3,&r4_5,&r6_7);
+  K_N(LEAFLUT[6],LEAFLUT[7],&r0_1,&r2_3,&r4_5,&r6_7);
   L_4_2(in+(*is)[8],in+(*is)[9],in+(*is)[10],in+(*is)[11],&r8_9,&r10_11,&r28_29,&r30_31);
   L_4_4(in+(*is)[12],in+(*is)[13],in+(*is)[14],in+(*is)[15],&r12_13,&r14_15,&r24_25,&r26_27);
-  K_N(_mm_set_ps(0.92387953251128673848313610506011,0.92387953251128673848313610506011,1,1),_mm_set_ps(0.38268343236508978177923268049199,-0.38268343236508978177923268049199,0,-0),&r0_1,&r4_5,&r8_9,&r12_13);
+  K_N(LEAFLUT[8],LEAFLUT[9],&r0_1,&r4_5,&r8_9,&r12_13);
   S_4(r0_1,r4_5,r8_9,r12_13,out0+0,out0+8,out0+16,out0+24);
-  K_N(_mm_set_ps(0.38268343236508983729038391174981,0.38268343236508983729038391174981,0.70710678118654757273731092936941,0.70710678118654757273731092936941),_mm_set_ps(0.92387953251128673848313610506011,-0.92387953251128673848313610506011,0.70710678118654746171500846685376,-0.70710678118654746171500846685376),&r2_3,&r6_7,&r10_11,&r14_15);
+  K_N(LEAFLUT[10],LEAFLUT[11],&r2_3,&r6_7,&r10_11,&r14_15);
   S_4(r2_3,r6_7,r10_11,r14_15,out0+4,out0+12,out0+20,out0+28);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0,-0),&r16_17,&r18_19,&r20_21,&r22_23);
+  K_N(LEAFLUT[6],LEAFLUT[7],&r16_17,&r18_19,&r20_21,&r22_23);
   S_4(r16_17,r18_19,r20_21,r22_23,out1+0,out1+4,out1+8,out1+12);
-  K_N(_mm_set_ps(0.70710678118654757273731092936941,0.70710678118654757273731092936941,1,1),_mm_set_ps(0.70710678118654746171500846685376,-0.70710678118654746171500846685376,0,-0),&r24_25,&r26_27,&r28_29,&r30_31);
+  K_N(LEAFLUT[6],LEAFLUT[7],&r24_25,&r26_27,&r28_29,&r30_31);
   S_4(r24_25,r26_27,r28_29,r30_31,out1+16,out1+20,out1+24,out1+28);
 
 	*is += 16;
diff --git a/src/patterns.c b/src/patterns.c
index 1ab593f..12d9a4c 100644
--- a/src/patterns.c
+++ b/src/patterns.c
@@ -66,7 +66,7 @@ void init_is(ffts_plan_t *p, int N, int leafN, int VL) {
 //	if(i % 16 == 15) printf("\n");
 //}
 
-	p->i0 = i0; p->i1 = i1; p->i2 = i2;
+	p->i0 = i0; p->i1 = i1;
 }
 
 void elaborate_offsets(ptrdiff_t *offsets, int leafN, int N, int ioffset, int ooffset, int stride, int even) {