ssedetect: Add runtime CPU detection

The current implementation can select the SSE support level during
compiletime only.

This commit adds functionality to automatically detect and switch
the SSE support level and automatically switch the Implementation
if the CPU does not support the required SSE level.

Change-Id: Iba74f8a6e4e921ff31e4bd9f0c7c881fe547423a

Transceiver52M/arm/convolve.c

@@ -58,6 +58,13 @@ static void neon_conv_cmplx_4n(float *x, float *h, float *y, int h_len, int len)
 }
 #endif
 
+/* API: Initalize convolve module */
+void convolve_init(void)
+{
+	/* Stub */
+	return;
+}
+
 /* API: Aligned complex-real */
 int convolve_real(float *x, int x_len,
 		  float *h, int h_len,

Transceiver52M/common/convert.h

@@ -3,5 +3,6 @@
 
 void convert_float_short(short *out, const float *in, float scale, int len);
 void convert_short_float(float *out, const short *in, int len);
+void convert_init(void);
 
 #endif /* _CONVERT_H_ */

Transceiver52M/common/convolve.h

@@ -27,4 +27,6 @@ int base_convolve_complex(const float *x, int x_len,
 			  int start, int len,
 			  int step, int offset);
 
+void convolve_init(void);
+
 #endif /* _CONVOLVE_H_ */

Transceiver52M/osmo-trx.cpp

@@ -32,6 +32,11 @@
 #include <Logger.h>
 #include <Configuration.h>
 
+extern "C" {
+#include "convolve.h"
+#include "convert.h"
+}
+
 /* Samples-per-symbol for downlink path
  *     4 - Uses precision modulator (more computation, less distortion)
  *     1 - Uses minimized modulator (less computation, more distortion)
@@ -422,6 +427,9 @@ int main(int argc, char *argv[])
 	RadioDevice::InterfaceType iface = RadioDevice::NORMAL;
 	struct trx_config config;
 
+	convolve_init();
+	convert_init();
+
 	handle_options(argc, argv, &config);
 
 	setup_signal_handlers();

Transceiver52M/x86/convert.c

@@ -25,6 +25,17 @@
 #include "config.h"
 #endif
 
+/* Architecture dependant function pointers */
+struct convert_cpu_context {
+	void (*convert_si16_ps_16n) (float *, const short *, int);
+	void (*convert_si16_ps) (float *, const short *, int);
+	void (*convert_scale_ps_si16_16n)(short *, const float *, float, int);
+	void (*convert_scale_ps_si16_8n)(short *, const float *, float, int);
+	void (*convert_scale_ps_si16)(short *, const float *, float, int);
+};
+
+static struct convert_cpu_context c;
+
 #ifdef HAVE_SSE3
 #include <xmmintrin.h>
 #include <emmintrin.h>
@@ -157,53 +168,61 @@ static void _sse_convert_scale_ps_si16_16n(short *restrict out,
 		_mm_storeu_si128((__m128i *) &out[16 * i + 8], m7);
 	}
 }
-#else /* HAVE_SSE3 */
+#endif
+
+__attribute__((optimize("no-tree-vectorize")))
 static void convert_scale_ps_si16(short *out, const float *in,
 				  float scale, int len)
 {
 	for (int i = 0; i < len; i++)
 		out[i] = in[i] * scale;
 }
-#endif
 
-#ifndef HAVE_SSE4_1
+__attribute__((optimize("no-tree-vectorize")))
 static void convert_si16_ps(float *out, const short *in, int len)
 {
 	for (int i = 0; i < len; i++)
 		out[i] = in[i];
 }
+
+void convert_init(void)
+{
+	c.convert_scale_ps_si16_16n = convert_scale_ps_si16;
+	c.convert_scale_ps_si16_8n = convert_scale_ps_si16;
+	c.convert_scale_ps_si16 = convert_scale_ps_si16;
+	c.convert_si16_ps_16n = convert_si16_ps;
+	c.convert_si16_ps = convert_si16_ps;
+
+#ifdef HAVE_SSE4_1
+	if (__builtin_cpu_supports("sse4.1")) {
+		c.convert_si16_ps_16n = &_sse_convert_si16_ps_16n;
+		c.convert_si16_ps = &_sse_convert_si16_ps;
+	}
 #endif
 
+#ifdef HAVE_SSE3
+	if (__builtin_cpu_supports("sse3")) {
+		c.convert_scale_ps_si16_16n = _sse_convert_scale_ps_si16_16n;
+		c.convert_scale_ps_si16_8n = _sse_convert_scale_ps_si16_8n;
+		c.convert_scale_ps_si16 = _sse_convert_scale_ps_si16;
+	}
+#endif
+}
+
 void convert_float_short(short *out, const float *in, float scale, int len)
 {
-	void (*conv_func)(short *, const float *, float, int);
-
-#ifdef HAVE_SSE3
 	if (!(len % 16))
-		conv_func = _sse_convert_scale_ps_si16_16n;
+		c.convert_scale_ps_si16_16n(out, in, scale, len);
 	else if (!(len % 8))
-		conv_func = _sse_convert_scale_ps_si16_8n;
+		c.convert_scale_ps_si16_8n(out, in, scale, len);
 	else
-		conv_func = _sse_convert_scale_ps_si16;
+		c.convert_scale_ps_si16(out, in, scale, len);
-#else
-	conv_func = convert_scale_ps_si16;
-#endif
-
-	conv_func(out, in, scale, len);
 }
 
 void convert_short_float(float *out, const short *in, int len)
 {
-	void (*conv_func) (float *, const short *, int);
-
-#ifdef HAVE_SSE4_1
 	if (!(len % 16))
-		conv_func = _sse_convert_si16_ps_16n;
+		c.convert_si16_ps_16n(out, in, len);
 	else
-		conv_func = _sse_convert_si16_ps;
+		c.convert_si16_ps(out, in, len);
-#else
-	conv_func = convert_si16_ps;
-#endif
-
-	conv_func(out, in, len);
 }

Transceiver52M/x86/convolve.c

@@ -26,6 +26,31 @@
 #include "config.h"
 #endif
 
+/* Architecture dependant function pointers */
+struct convolve_cpu_context {
+	void (*conv_cmplx_4n) (const float *, int, const float *, int, float *,
+			       int, int, int, int, int);
+	void (*conv_cmplx_8n) (const float *, int, const float *, int, float *,
+			       int, int, int, int, int);
+	void (*conv_cmplx) (const float *, int, const float *, int, float *,
+			    int, int, int, int, int);
+	void (*conv_real4) (const float *, int, const float *, int, float *,
+			    int, int, int, int, int);
+	void (*conv_real8) (const float *, int, const float *, int, float *,
+			    int, int, int, int, int);
+	void (*conv_real12) (const float *, int, const float *, int, float *,
+			     int, int, int, int, int);
+	void (*conv_real16) (const float *, int, const float *, int, float *,
+			     int, int, int, int, int);
+	void (*conv_real20) (const float *, int, const float *, int, float *,
+			     int, int, int, int, int);
+	void (*conv_real4n) (const float *, int, const float *, int, float *,
+			     int, int, int, int, int);
+	void (*conv_real) (const float *, int, const float *, int, float *, int,
+			   int, int, int, int);
+};
+static struct convolve_cpu_context c;
+
 /* Forward declarations from base implementation */
 int _base_convolve_real(const float *x, int x_len,
 			const float *h, int h_len,
@@ -565,45 +590,77 @@ static void sse_conv_cmplx_8n(const float *x, int x_len,
 }
 #endif
 
+/* API: Initalize convolve module */
+void convolve_init(void)
+{
+	c.conv_cmplx_4n = (void *)_base_convolve_complex;
+	c.conv_cmplx_8n = (void *)_base_convolve_complex;
+	c.conv_cmplx = (void *)_base_convolve_complex;
+	c.conv_real4 = (void *)_base_convolve_real;
+	c.conv_real8 = (void *)_base_convolve_real;
+	c.conv_real12 = (void *)_base_convolve_real;
+	c.conv_real16 = (void *)_base_convolve_real;
+	c.conv_real20 = (void *)_base_convolve_real;
+	c.conv_real4n = (void *)_base_convolve_real;
+	c.conv_real = (void *)_base_convolve_real;
+
+#ifdef HAVE_SSE3
+	if (__builtin_cpu_supports("sse3")) {
+		c.conv_cmplx_4n = sse_conv_cmplx_4n;
+		c.conv_cmplx_8n = sse_conv_cmplx_8n;
+		c.conv_real4 = sse_conv_real4;
+		c.conv_real8 = sse_conv_real8;
+		c.conv_real12 = sse_conv_real12;
+		c.conv_real16 = sse_conv_real16;
+		c.conv_real20 = sse_conv_real20;
+		c.conv_real4n = sse_conv_real4n;
+	}
+#endif
+}
+
 /* API: Aligned complex-real */
 int convolve_real(const float *x, int x_len,
 		  const float *h, int h_len,
 		  float *y, int y_len, int start, int len, int step, int offset)
 {
-	void (*conv_func) (const float *, int, const float *, int, float *, int,
-			   int, int, int, int) = (void *)_base_convolve_real;
-
 	if (bounds_check(x_len, h_len, y_len, start, len, step) < 0)
 		return -1;
 
 	memset(y, 0, len * 2 * sizeof(float));
 
-#ifdef HAVE_SSE3
 	if (step <= 4) {
 		switch (h_len) {
 		case 4:
-			conv_func = sse_conv_real4;
+			c.conv_real4(x, x_len, h, h_len, y, y_len, start, len,
+				     step, offset);
 			break;
 		case 8:
-			conv_func = sse_conv_real8;
+			c.conv_real8(x, x_len, h, h_len, y, y_len, start, len,
+				     step, offset);
 			break;
 		case 12:
-			conv_func = sse_conv_real12;
+			c.conv_real12(x, x_len, h, h_len, y, y_len, start, len,
+				      step, offset);
 			break;
 		case 16:
-			conv_func = sse_conv_real16;
+			c.conv_real16(x, x_len, h, h_len, y, y_len, start, len,
+				      step, offset);
 			break;
 		case 20:
-			conv_func = sse_conv_real20;
+			c.conv_real20(x, x_len, h, h_len, y, y_len, start, len,
+				      step, offset);
 			break;
 		default:
 			if (!(h_len % 4))
-				conv_func = sse_conv_real4n;
+				c.conv_real4n(x, x_len, h, h_len, y, y_len,
+					      start, len, step, offset);
+			else
+				c.conv_real(x, x_len, h, h_len, y, y_len, start,
+					    len, step, offset);
 		}
-	}
+	} else
-#endif
+		c.conv_real(x, x_len, h, h_len, y, y_len, start, len, step,
-
+			    offset);
-	conv_func(x, x_len, h, h_len, y, y_len, start, len, step, offset);
 
 	return len;
 }
@@ -614,25 +671,24 @@ int convolve_complex(const float *x, int x_len,
 		     float *y, int y_len,
 		     int start, int len, int step, int offset)
 {
-	void (*conv_func) (const float *, int, const float *, int, float *, int,
-			   int, int, int, int) =
-	    (void *)_base_convolve_complex;
-
 	if (bounds_check(x_len, h_len, y_len, start, len, step) < 0)
 		return -1;
 
 	memset(y, 0, len * 2 * sizeof(float));
 
-#ifdef HAVE_SSE3
 	if (step <= 4) {
 		if (!(h_len % 8))
-			conv_func = sse_conv_cmplx_8n;
+			c.conv_cmplx_8n(x, x_len, h, h_len, y, y_len, start,
+					len, step, offset);
 		else if (!(h_len % 4))
-			conv_func = sse_conv_cmplx_4n;
+			c.conv_cmplx_4n(x, x_len, h, h_len, y, y_len, start,
-	}
+					len, step, offset);
-#endif
+		else
-
+			c.conv_cmplx(x, x_len, h, h_len, y, y_len, start, len,
-	conv_func(x, x_len, h, h_len, y, y_len, start, len, step, offset);
+				     step, offset);
+	} else
+		c.conv_cmplx(x, x_len, h, h_len, y, y_len, start, len, step,
+			     offset);
 
 	return len;
 }