Implement fast SSE2 lrint and lrintf, closes libsndfile#660

lrint() and lrintf() are ridiculously slow when libsndfile is built
using Visual C++.

On x86 and AMD64 platforms this patch adds SSE2 optimized versions.

It is safe to assume that non-SSE2 processors are no longer available to
avoid unnecessary checks.

NEWS

@@ -1,4 +1,7 @@
 Unreleased
+  * Implement fast SSE2 optimized psf_lrintf() and psf_lrintf() functions to
+    improve perfomance when libsndfile is built using Visual C++ (especially)
+    and other compilers on x86 and AMD64 platforms.
   * Documentation:
     * Move site to new URL: http://libsndfile.github.io/libsndfile/
     * Convert documentation pages from HTML to Markdown

cmake/SndFileChecks.cmake

@@ -67,6 +67,7 @@ check_include_file (stdint.h		HAVE_STDINT_H)
 check_include_file (sys/time.h		HAVE_SYS_TIME_H)
 check_include_file (sys/types.h		HAVE_SYS_TYPES_H)
 check_include_file (unistd.h		HAVE_UNISTD_H)
+check_include_file (immintrin.h		HAVE_IMMINTRIN_H)
 
 # Never checked
 # check_include_file (stdlib.h		HAVE_STDLIB_H)

configure.ac

@@ -100,6 +100,7 @@ AC_CHECK_HEADERS([endian.h])
 AC_CHECK_HEADERS([byteswap.h])
 AC_CHECK_HEADERS([locale.h])
 AC_CHECK_HEADERS([sys/time.h])
+AC_CHECK_HEADERS([immintrin.h])
 
 AC_HEADER_SYS_WAIT
 

src/alaw.c

@@ -340,9 +340,9 @@ static inline void
 f2alaw_array (const float *ptr, int count, unsigned char *buffer, float normfact)
 {	while (--count >= 0)
 	{	if (ptr [count] >= 0)
-			buffer [count] = alaw_encode [lrintf (normfact * ptr [count])] ;
+			buffer [count] = alaw_encode [psf_lrintf (normfact * ptr [count])] ;
 		else
-			buffer [count] = 0x7F & alaw_encode [- lrintf (normfact * ptr [count])] ;
+			buffer [count] = 0x7F & alaw_encode [- psf_lrintf (normfact * ptr [count])] ;
 		} ;
 } /* f2alaw_array */
 
@@ -352,9 +352,9 @@ d2alaw_array (const double *ptr, int count, unsigned char *buffer, double normfa
 	{	if (!isfinite (ptr [count]))
 			buffer [count] = 0 ;
 		else if (ptr [count] >= 0)
-			buffer [count] = alaw_encode [lrint (normfact * ptr [count])] ;
+			buffer [count] = alaw_encode [psf_lrint (normfact * ptr [count])] ;
 		else
-			buffer [count] = 0x7F & alaw_encode [- lrint (normfact * ptr [count])] ;
+			buffer [count] = 0x7F & alaw_encode [- psf_lrint (normfact * ptr [count])] ;
 		} ;
 } /* d2alaw_array */
 

src/caf.c

@@ -370,7 +370,7 @@ caf_read_header (SF_PRIVATE *psf)
 		return SFE_MALFORMED_FILE ;
 		} ;
 
-	psf->sf.samplerate = lrint (srate) ;
+	psf->sf.samplerate = psf_lrint (srate) ;
 
 	psf_binheader_readf (psf, "mE44444", &desc.fmt_id, &desc.fmt_flags, &desc.pkt_bytes, &desc.frames_per_packet,
 			&desc.channels_per_frame, &desc.bits_per_chan) ;

src/common.c

@@ -1606,7 +1606,7 @@ psf_f2s_array (const float *src, short *dest, int count, int normalize)
 
 	normfact = normalize ? (1.0 * 0x7FFF) : 1.0 ;
 	while (--count >= 0)
-		dest [count] = lrintf (src [count] * normfact) ;
+		dest [count] = psf_lrintf (src [count] * normfact) ;
 
 	return ;
 } /* psf_f2s_array */
@@ -1628,7 +1628,7 @@ psf_f2s_clip_array (const float *src, short *dest, int count, int normalize)
 			continue ;
 			} ;
 
-		dest [count] = lrintf (scaled_value) ;
+		dest [count] = psf_lrintf (scaled_value) ;
 		} ;
 
 	return ;
@@ -1640,7 +1640,7 @@ psf_d2s_array (const double *src, short *dest, int count, int normalize)
 
 	normfact = normalize ? (1.0 * 0x7FFF) : 1.0 ;
 	while (--count >= 0)
-		dest [count] = lrint (src [count] * normfact) ;
+		dest [count] = psf_lrint (src [count] * normfact) ;
 
 	return ;
 } /* psf_f2s_array */
@@ -1662,7 +1662,7 @@ psf_d2s_clip_array (const double *src, short *dest, int count, int normalize)
 			continue ;
 			} ;
 
-		dest [count] = lrint (scaled_value) ;
+		dest [count] = psf_lrint (scaled_value) ;
 		} ;
 
 	return ;
@@ -1675,7 +1675,7 @@ psf_f2i_array (const float *src, int *dest, int count, int normalize)
 
 	normfact = normalize ? (1.0 * 0x7FFFFFFF) : 1.0 ;
 	while (--count >= 0)
-		dest [count] = lrintf (src [count] * normfact) ;
+		dest [count] = psf_lrintf (src [count] * normfact) ;
 
 	return ;
 } /* psf_f2i_array */
@@ -1697,7 +1697,7 @@ psf_f2i_clip_array (const float *src, int *dest, int count, int normalize)
 			continue ;
 			} ;
 
-		dest [count] = lrintf (scaled_value) ;
+		dest [count] = psf_lrintf (scaled_value) ;
 		} ;
 
 	return ;
@@ -1709,7 +1709,7 @@ psf_d2i_array (const double *src, int *dest, int count, int normalize)
 
 	normfact = normalize ? (1.0 * 0x7FFFFFFF) : 1.0 ;
 	while (--count >= 0)
-		dest [count] = lrint (src [count] * normfact) ;
+		dest [count] = psf_lrint (src [count] * normfact) ;
 
 	return ;
 } /* psf_f2i_array */
@@ -1731,7 +1731,7 @@ psf_d2i_clip_array (const double *src, int *dest, int count, int normalize)
 			continue ;
 			} ;
 
-		dest [count] = lrint (scaled_value) ;
+		dest [count] = psf_lrint (scaled_value) ;
 		} ;
 
 	return ;

src/common.h

@@ -37,6 +37,10 @@
 #include "sndfile.h"
 #endif
 
+#ifdef USE_SSE2
+#include <immintrin.h>
+#endif
+
 #ifdef __cplusplus
 #error "This code is not designed to be compiled with a C++ compiler."
 #endif
@@ -546,6 +550,8 @@ typedef struct sf_private_tag
 	SF_CHUNK_ITERATOR *	(*next_chunk_iterator)	(struct sf_private_tag*, SF_CHUNK_ITERATOR * iterator) ;
 	int					(*get_chunk_size)	(struct sf_private_tag*, const SF_CHUNK_ITERATOR * iterator, SF_CHUNK_INFO * chunk_info) ;
 	int					(*get_chunk_data)	(struct sf_private_tag*, const SF_CHUNK_ITERATOR * iterator, SF_CHUNK_INFO * chunk_info) ;
+
+	int cpu_flags ;
 } SF_PRIVATE ;
 
 
@@ -998,6 +1004,28 @@ psf_strlcpy (char *dest, size_t n, const char *src)
 	dest [n - 1] = 0 ;
 } /* psf_strlcpy */
 
+/*------------------------------------------------------------------------------------
+** SIMD optimized math functions.
+*/
+
+static inline int psf_lrintf (float x)
+{
+	#ifdef USE_SSE2
+ 		return _mm_cvtss_si32 (_mm_load_ss (&x)) ;
+	#else
+		return lrintf (x) ;
+	#endif
+} /* psf_lrintf */
+
+static inline int psf_lrint (double x)
+{
+	#ifdef USE_SSE2
+ 		return _mm_cvtsd_si32 (_mm_load_sd (&x)) ;
+	#else
+		return lrint (x) ;
+	#endif
+} /* psf_lrintf */
+
 /*------------------------------------------------------------------------------------
 ** Other helper functions.
 */

src/config.h.cmake

@@ -172,6 +172,9 @@
 /* Define to 1 if you have the <unistd.h> header file. */
 #cmakedefine01 HAVE_UNISTD_H
 
+/* Define to 1 if you have the <immintrin.h> header file. */
+#cmakedefine01 HAVE_IMMINTRIN_H
+
 /* Define to 1 if you have the `vsnprintf' function. */
 #cmakedefine01 HAVE_VSNPRINTF
 

src/double64.c

@@ -368,7 +368,7 @@ double64_be_write (double in, unsigned char *out)
 	out [1] |= (exponent << 4) & 0xF0 ;
 
 	in *= 0x20000000 ;
-	mantissa = lrint (floor (in)) ;
+	mantissa = psf_lrint (floor (in)) ;
 
 	out [1] |= (mantissa >> 24) & 0xF ;
 	out [2] = (mantissa >> 16) & 0xFF ;
@@ -377,7 +377,7 @@ double64_be_write (double in, unsigned char *out)
 
 	in = fmod (in, 1.0) ;
 	in *= 0x1000000 ;
-	mantissa = lrint (floor (in)) ;
+	mantissa = psf_lrint (floor (in)) ;
 
 	out [5] = (mantissa >> 16) & 0xFF ;
 	out [6] = (mantissa >> 8) & 0xFF ;
@@ -408,7 +408,7 @@ double64_le_write (double in, unsigned char *out)
 	out [6] |= (exponent << 4) & 0xF0 ;
 
 	in *= 0x20000000 ;
-	mantissa = lrint (floor (in)) ;
+	mantissa = psf_lrint (floor (in)) ;
 
 	out [6] |= (mantissa >> 24) & 0xF ;
 	out [5] = (mantissa >> 16) & 0xFF ;
@@ -417,7 +417,7 @@ double64_le_write (double in, unsigned char *out)
 
 	in = fmod (in, 1.0) ;
 	in *= 0x1000000 ;
-	mantissa = lrint (floor (in)) ;
+	mantissa = psf_lrint (floor (in)) ;
 
 	out [2] = (mantissa >> 16) & 0xFF ;
 	out [1] = (mantissa >> 8) & 0xFF ;
@@ -487,7 +487,7 @@ double64_get_capability	(SF_PRIVATE *psf)
 static void
 d2s_array (const double *src, int count, short *dest, double scale)
 {	while (--count >= 0)
-	{	dest [count] = lrint (scale * src [count]) ;
+	{	dest [count] = psf_lrint (scale * src [count]) ;
 		} ;
 } /* d2s_array */
 
@@ -501,14 +501,14 @@ d2s_clip_array (const double *src, int count, short *dest, double scale)
 		else if (CPU_CLIPS_NEGATIVE == 0 && tmp < -32768.0)
 			dest [count] = SHRT_MIN ;
 		else
-			dest [count] = lrint (tmp) ;
+			dest [count] = psf_lrint (tmp) ;
 		} ;
 } /* d2s_clip_array */
 
 static void
 d2i_array (const double *src, int count, int *dest, double scale)
 {	while (--count >= 0)
-	{	dest [count] = lrint (scale * src [count]) ;
+	{	dest [count] = psf_lrint (scale * src [count]) ;
 		} ;
 } /* d2i_array */
 
@@ -522,7 +522,7 @@ d2i_clip_array (const double *src, int count, int *dest, double scale)
 		else if (CPU_CLIPS_NEGATIVE == 0 && tmp < (-1.0 * INT_MAX))
 			dest [count] = INT_MIN ;
 		else
-			dest [count] = lrint (tmp) ;
+			dest [count] = psf_lrint (tmp) ;
 		} ;
 } /* d2i_clip_array */
 

src/dwvw.c

@@ -628,7 +628,7 @@ dwvw_write_f (SF_PRIVATE *psf, const float *ptr, sf_count_t len)
 	while (len > 0)
 	{	writecount = (len >= bufferlen) ? bufferlen : len ;
 		for (k = 0 ; k < writecount ; k++)
-			iptr [k] = lrintf (normfact * ptr [total + k]) ;
+			iptr [k] = psf_lrintf (normfact * ptr [total + k]) ;
 		count = dwvw_encode_data (psf, pdwvw, iptr, writecount) ;
 
 		total += count ;
@@ -660,7 +660,7 @@ dwvw_write_d (SF_PRIVATE *psf, const double *ptr, sf_count_t len)
 	while (len > 0)
 	{	writecount = (len >= bufferlen) ? bufferlen : len ;
 		for (k = 0 ; k < writecount ; k++)
-			iptr [k] = lrint (normfact * ptr [total + k]) ;
+			iptr [k] = psf_lrint (normfact * ptr [total + k]) ;
 		count = dwvw_encode_data (psf, pdwvw, iptr, writecount) ;
 
 		total += count ;

src/flac.c

@@ -881,7 +881,7 @@ flac_command (SF_PRIVATE * psf, int command, void * data, int datasize)
 			*/
 			quality = (*((double *) data)) * 8.0 ;
 			/* Clip range. */
-			pflac->compression = lrint (SF_MAX (0.0, SF_MIN (8.0, quality))) ;
+			pflac->compression = psf_lrint (SF_MAX (0.0, SF_MIN (8.0, quality))) ;
 
 			psf_log_printf (psf, "%s : Setting SFC_SET_COMPRESSION_LEVEL to %u.\n", __func__, pflac->compression) ;
 
@@ -1186,7 +1186,7 @@ f2flac8_clip_array (const float *src, int32_t *dest, int count, int normalize)
 		{	dest [count] = -0x80 ;
 			continue ;
 			} ;
-		dest [count] = lrintf (scaled_value) ;
+		dest [count] = psf_lrintf (scaled_value) ;
 		} ;
 
 	return ;
@@ -1208,7 +1208,7 @@ f2flac16_clip_array (const float *src, int32_t *dest, int count, int normalize)
 		{	dest [count] = -0x8000 ;
 			continue ;
 			} ;
-		dest [count] = lrintf (scaled_value) ;
+		dest [count] = psf_lrintf (scaled_value) ;
 		} ;
 } /* f2flac16_clip_array */
 
@@ -1229,7 +1229,7 @@ f2flac24_clip_array (const float *src, int32_t *dest, int count, int normalize)
 		{	dest [count] = -0x800000 ;
 			continue ;
 			}
-		dest [count] = lrintf (scaled_value) ;
+		dest [count] = psf_lrintf (scaled_value) ;
 		} ;
 
 	return ;
@@ -1240,23 +1240,23 @@ f2flac8_array (const float *src, int32_t *dest, int count, int normalize)
 {	float normfact = normalize ? (1.0 * 0x7F) : 1.0 ;
 
 	while (--count >= 0)
-		dest [count] = lrintf (src [count] * normfact) ;
+		dest [count] = psf_lrintf (src [count] * normfact) ;
 } /* f2flac8_array */
 
 static void
 f2flac16_array (const float *src, int32_t *dest, int count, int normalize)
 {	float normfact = normalize ? (1.0 * 0x7FFF) : 1.0 ;
 
 	while (--count >= 0)
-		dest [count] = lrintf (src [count] * normfact) ;
+		dest [count] = psf_lrintf (src [count] * normfact) ;
 } /* f2flac16_array */
 
 static void
 f2flac24_array (const float *src, int32_t *dest, int count, int normalize)
 {	float normfact = normalize ? (1.0 * 0x7FFFFF) : 1.0 ;
 
 	while (--count >= 0)
-		dest [count] = lrintf (src [count] * normfact) ;
+		dest [count] = psf_lrintf (src [count] * normfact) ;
 } /* f2flac24_array */
 
 static sf_count_t
@@ -1317,7 +1317,7 @@ d2flac8_clip_array (const double *src, int32_t *dest, int count, int normalize)
 		{	dest [count] = -0x80 ;
 			continue ;
 			} ;
-		dest [count] = lrint (scaled_value) ;
+		dest [count] = psf_lrint (scaled_value) ;
 		} ;
 
 	return ;
@@ -1339,7 +1339,7 @@ d2flac16_clip_array (const double *src, int32_t *dest, int count, int normalize)
 		{	dest [count] = -0x8000 ;
 			continue ;
 			} ;
-		dest [count] = lrint (scaled_value) ;
+		dest [count] = psf_lrint (scaled_value) ;
 		} ;
 
 	return ;
@@ -1361,7 +1361,7 @@ d2flac24_clip_array (const double *src, int32_t *dest, int count, int normalize)
 		{	dest [count] = -0x800000 ;
 			continue ;
 			} ;
-		dest [count] = lrint (scaled_value) ;
+		dest [count] = psf_lrint (scaled_value) ;
 		} ;
 
 	return ;
@@ -1372,23 +1372,23 @@ d2flac8_array (const double *src, int32_t *dest, int count, int normalize)
 {	double normfact = normalize ? (1.0 * 0x7F) : 1.0 ;
 
 	while (--count >= 0)
-		dest [count] = lrint (src [count] * normfact) ;
+		dest [count] = psf_lrint (src [count] * normfact) ;
 } /* d2flac8_array */
 
 static void
 d2flac16_array (const double *src, int32_t *dest, int count, int normalize)
 {	double normfact = normalize ? (1.0 * 0x7FFF) : 1.0 ;
 
 	while (--count >= 0)
-		dest [count] = lrint (src [count] * normfact) ;
+		dest [count] = psf_lrint (src [count] * normfact) ;
 } /* d2flac16_array */
 
 static void
 d2flac24_array (const double *src, int32_t *dest, int count, int normalize)
 {	double normfact = normalize ? (1.0 * 0x7FFFFF) : 1.0 ;
 
 	while (--count >= 0)
-		dest [count] = lrint (src [count] * normfact) ;
+		dest [count] = psf_lrint (src [count] * normfact) ;
 } /* d2flac24_array */
 
 static sf_count_t