diff --git a/_kiss_fft_guts.h b/_kiss_fft_guts.h
index 7bf5762..33eea21 100644
--- a/_kiss_fft_guts.h
+++ b/_kiss_fft_guts.h
@@ -119,6 +119,52 @@ struct kiss_fft_state{
     }while(0)
 
 
+// Normally same as kiss_fft_scalar, but for USE_SIMD this is only a single number
+#ifndef kiss_fft_scalar_one
+#define kiss_fft_scalar_one kiss_fft_scalar
+#endif
+
+
+// If kiss_fft_suffix is not provided, determine it automatically based on kiss_fft_scalar_one
+#if !defined (kiss_fft_suffix) && !defined (kiss_fft_suffix_empty)
+#ifdef FIXED_POINT
+#undef kiss_fft_suffix
+#define kiss_fft_suffix_empty 1
+#endif
+
+#define KISS_X_float_X 1
+#define KISS_X__X 2
+#define KISS_X_long_X 3
+#define KISS_concat2(a, b, c) a##b##c
+#define KISS_concat(a, b, c) KISS_concat2(a, b, c)
+
+#define double // This is because otherwise KISS_concat(KISS_X_, long double, _X) would produce two tokens which would break the preprocessor comparison
+
+#if KISS_concat(KISS_X_, kiss_fft_scalar_one, _X) == KISS_X_float_X // float
+#define kiss_fft_suffix f
+#elif KISS_concat(KISS_X_, kiss_fft_scalar_one, _X) == KISS_X__X // double
+#undef kiss_fft_suffix
+#define kiss_fft_suffix_empty 1
+#elif KISS_concat(KISS_X_, kiss_fft_scalar_one, _X) == KISS_X_long_X // long double
+#define kiss_fft_suffix l
+#else
+#undef kiss_fft_suffix
+#define kiss_fft_suffix_empty 1
+#warning "Unknown kiss_fft_scalar type"
+#endif
+
+#undef double
+#endif
+
+
+#ifdef kiss_fft_suffix_empty
+#define KISS_ADD_SUFFIX(x) x
+#else
+#define KISS_ADD_SUFFIX2(x, y) x##y
+#define KISS_ADD_SUFFIX1(x, y) KISS_ADD_SUFFIX2(x, y)
+#define KISS_ADD_SUFFIX(x) KISS_ADD_SUFFIX1(x, kiss_fft_suffix)
+#endif
+
 #ifdef FIXED_POINT
 #  define KISS_FFT_COS(phase)  floor(.5+SAMP_MAX * cos (phase))
 #  define KISS_FFT_SIN(phase)  floor(.5+SAMP_MAX * sin (phase))
@@ -128,9 +174,9 @@ struct kiss_fft_state{
 #  define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) )
 #  define HALF_OF(x) ((x)*_mm_set1_ps(.5))
 #else
-#  define KISS_FFT_COS(phase) (kiss_fft_scalar) cos(phase)
-#  define KISS_FFT_SIN(phase) (kiss_fft_scalar) sin(phase)
-#  define HALF_OF(x) ((x)*.5)
+#  define KISS_FFT_COS(phase) (kiss_fft_scalar) KISS_ADD_SUFFIX(cos)(phase)
+#  define KISS_FFT_SIN(phase) (kiss_fft_scalar) KISS_ADD_SUFFIX(sin)(phase)
+#  define HALF_OF(x) ((x)*KISS_ADD_SUFFIX(.5))
 #endif
 
 #define  kf_cexp(x,phase) \
diff --git a/kiss_fft.c b/kiss_fft.c
index 5614838..058683d 100644
--- a/kiss_fft.c
+++ b/kiss_fft.c
@@ -349,8 +349,8 @@ kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem
         st->inverse = inverse_fft;
 
         for (i=0;i<nfft;++i) {
-            const double pi=3.141592653589793238462643383279502884197169399375105820974944;
-            double phase = -2*pi*i / nfft;
+            const kiss_fft_scalar_one pi=KISS_ADD_SUFFIX(3.141592653589793238462643383279502884197169399375105820974944);
+            kiss_fft_scalar_one phase = -2*pi*i / nfft;
             if (st->inverse)
                 phase *= -1;
             kf_cexp(st->twiddles+i, phase );
diff --git a/kiss_fft.h b/kiss_fft.h
index ed6e24a..4bf5c7e 100644
--- a/kiss_fft.h
+++ b/kiss_fft.h
@@ -35,6 +35,8 @@ extern "C" {
 #ifdef USE_SIMD
 # include <xmmintrin.h>
 # define kiss_fft_scalar __m128
+# define kiss_fft_scalar_one double
+//# define kiss_fft_scalar_one float
 # ifndef KISS_FFT_MALLOC
 #  define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes,16)
 # endif
@@ -55,8 +57,10 @@ extern "C" {
 #include <stdint.h>
 # if (FIXED_POINT == 32)
 #  define kiss_fft_scalar int32_t
+#  define kiss_fft_scalar_one double
 # else	
 #  define kiss_fft_scalar int16_t
+#  define kiss_fft_scalar_one double
 # endif
 #else
 # ifndef kiss_fft_scalar
diff --git a/tools/kiss_fftr.c b/tools/kiss_fftr.c
index 8102132..9c380ff 100644
--- a/tools/kiss_fftr.c
+++ b/tools/kiss_fftr.c
@@ -49,8 +49,8 @@ kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem,size_t * lenme
     kiss_fft_alloc(nfft, inverse_fft, st->substate, &subsize);
 
     for (i = 0; i < nfft/2; ++i) {
-        double phase =
-            -3.14159265358979323846264338327 * ((double) (i+1) / nfft + .5);
+        kiss_fft_scalar_one phase =
+            KISS_ADD_SUFFIX(-3.14159265358979323846264338327) * ((kiss_fft_scalar_one) (i+1) / nfft + .5);
         if (inverse_fft)
             phase *= -1;
         kf_cexp (st->super_twiddles+i,phase);