Merge pull request #182 from atong-tcs/sse2

-DWITH_SSE2 now does runtime detection

CMakeLists.txt

@@ -76,7 +76,7 @@ if(CMAKE_COMPILER_IS_GNUCC)
 	if(CMAKE_BUILD_TYPE STREQUAL "Release")
 		set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -O2")
 	endif()
-	if(WITH_SSE2)
+	if(WITH_SSE2_TARGET)
 		set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -msse2")
 	endif()
 endif()

client/X11/xfreerdp.c

@@ -570,7 +570,7 @@ boolean xf_post_connect(freerdp* instance)
 		xfi->primary_buffer = (uint8*) xzalloc(xfi->width * xfi->height * xfi->bpp);
 
 		if (instance->settings->rfx_codec)
-			xfi->rfx_context = (void*) rfx_context_new();
+			xfi->rfx_context = (void*) rfx_context_new(instance->settings);
 
 		if (instance->settings->ns_codec)
 			xfi->nsc_context = (void*) nsc_context_new();
@@ -675,6 +675,35 @@ boolean xf_verify_certificate(freerdp* instance, char* subject, char* issuer, ch
 }
 
 
+void cpuid(unsigned info, unsigned *eax, unsigned *ebx, unsigned *ecx, unsigned *edx)
+{
+	*eax = info;
+	__asm volatile
+		("mov %%ebx, %%edi;" /* 32bit PIC: don't clobber ebx */
+		 "cpuid;"
+		 "mov %%ebx, %%esi;"
+		 "mov %%edi, %%ebx;"
+		 :"+a" (*eax), "=S" (*ebx), "=c" (*ecx), "=d" (*edx)
+		 : :"edi");
+}
+ 
+uint32 xf_detect_cpu()
+{
+	unsigned int eax, ebx, ecx, edx;
+	uint32 cpu_opt = 0;
+
+	cpuid(1, &eax, &ebx, &ecx, &edx);
+
+	if (edx & (1<<26)) 
+	{
+		DEBUG("SSE2 detected");
+		cpu_opt |= CPU_SSE2;
+	}
+
+	return cpu_opt;
+}
+
+
 int xf_process_client_args(rdpSettings* settings, const char* opt, const char* val, void* user_data)
 {
 	int argc = 0;
@@ -971,6 +1000,11 @@ int main(int argc, char* argv[])
 	instance->context->argv = argv;
 	instance->settings->sw_gdi = False;
 
+#ifdef WITH_SSE2
+	/* detect only if needed */
+	instance->settings->cpu_opt = xf_detect_cpu();
+#endif
+
 	data = (struct thread_data*) xzalloc(sizeof(struct thread_data));
 	data->instance = instance;
 

cmake/ConfigOptions.cmake

@@ -12,6 +12,9 @@ option(WITH_DEBUG_RFX "Print RemoteFX debug messages." OFF)
 option(WITH_DEBUG_X11 "Print X11 Client debug messages" OFF)
 option(WITH_DEBUG_RAIL "Print RemoteApp debug messages" OFF)
 option(WITH_DEBUG_XV "Print XVideo debug messages" OFF)
+option(WITH_DEBUG_SCARD "Print smartcard debug messages" OFF)
+option(WITH_DEBUG_ORDERS "Print drawing orders debug messages" OFF)
 option(WITH_MANPAGES "Generate manpages." ON)
 option(WITH_PROFILER "Compile profiler." OFF)
 option(WITH_SSE2 "Use SSE2 optimization." OFF)
+option(WITH_SSE2_TARGET "Allow compiler to generate SSE2 instructions." OFF)

config.h.in

@@ -31,9 +31,11 @@
 #cmakedefine WITH_DEBUG_RFX
 #cmakedefine WITH_PROFILER
 #cmakedefine WITH_SSE2
+#cmakedefine WITH_SSE2_TARGET
 #cmakedefine WITH_DEBUG_X11
 #cmakedefine WITH_DEBUG_RAIL
 #cmakedefine WITH_DEBUG_XV
 #cmakedefine WITH_DEBUG_SCARD
+#cmakedefine WITH_DEBUG_ORDERS
 
 #endif

cunit/test_librfx.c

@@ -267,7 +267,7 @@ void test_dwt(void)
 {
 	RFX_CONTEXT* context;
 
-	context = rfx_context_new();
+	context = rfx_context_new(NULL);
 	rfx_dwt_2d_decode(buffer, context->priv->dwt_buffer);
 	//dump_buffer(buffer, 4096);
 	rfx_context_free(context);
@@ -304,7 +304,7 @@ void test_decode(void)
 	stream_write(s, cr_data, sizeof(cr_data));
 	stream_set_pos(s, 0);
 
-	context = rfx_context_new();
+	context = rfx_context_new(NULL);
 	context->mode = RLGR3;
 	rfx_context_set_pixel_format(context, RFX_PIXEL_FORMAT_RGB);
 	rfx_decode_rgb(context, s,
@@ -334,7 +334,7 @@ void test_encode(void)
 	enc_stream = stream_new(65536);
 	stream_clear(enc_stream);
 
-	context = rfx_context_new();
+	context = rfx_context_new(NULL);
 	context->mode = RLGR3;
 	rfx_context_set_pixel_format(context, RFX_PIXEL_FORMAT_RGB);
 
@@ -378,7 +378,7 @@ void test_message(void)
 	s = stream_new(65536);
 	stream_clear(s);
 
-	context = rfx_context_new();
+	context = rfx_context_new(NULL);
 	context->mode = RLGR3;
 	context->width = 800;
 	context->height = 600;

include/freerdp/codec/rfx.h

@@ -125,7 +125,7 @@ struct _RFX_CONTEXT
 };
 typedef struct _RFX_CONTEXT RFX_CONTEXT;
 
-FREERDP_API RFX_CONTEXT* rfx_context_new(void);
+FREERDP_API RFX_CONTEXT* rfx_context_new(rdpSettings* settings);
 FREERDP_API void rfx_context_free(RFX_CONTEXT* context);
 FREERDP_API void rfx_context_set_pixel_format(RFX_CONTEXT* context, RFX_PIXEL_FORMAT pixel_format);
 FREERDP_API void rfx_context_reset(RFX_CONTEXT* context);

include/freerdp/settings.h

@@ -90,6 +90,9 @@
 #define GLYPH_SUPPORT_FULL			0x0002
 #define GLYPH_SUPPORT_ENCODE			0x0003
 
+/* CPU Optimization flags */
+#define CPU_SSE2				0x1
+
 /* SYSTEM_TIME */
 typedef struct
 {
@@ -325,6 +328,8 @@ struct rdp_settings
 	boolean rail_langbar_supported;
 
 	boolean mouse_motion;
+
+	uint32 cpu_opt;
 };
 typedef struct rdp_settings rdpSettings;
 

libfreerdp-codec/CMakeLists.txt

@@ -46,6 +46,7 @@ if(WITH_SSE2)
 	rfx_sse2.c
 	rfx_sse2.h
 )
+	set_property(SOURCE rfx_sse2.c PROPERTY COMPILE_FLAGS "-msse2")
 endif()
 
 add_library(freerdp-codec ${FREERDP_CODEC_SRCS})

libfreerdp-codec/rfx.c

@@ -128,7 +128,7 @@ static void rfx_profiler_print(RFX_CONTEXT* context)
 	PROFILER_PRINT_FOOTER;
 }
 
-RFX_CONTEXT* rfx_context_new(void)
+RFX_CONTEXT* rfx_context_new(rdpSettings* settings)
 {
 	RFX_CONTEXT* context;
 
@@ -157,8 +157,9 @@ RFX_CONTEXT* rfx_context_new(void)
 	context->dwt_2d_decode = rfx_dwt_2d_decode;
 	context->dwt_2d_encode = rfx_dwt_2d_encode;
 
-	/* detect and enable SIMD CPU acceleration */
-	RFX_INIT_SIMD(context);
+	/* enable SIMD CPU acceleration if detected */
+	if (settings && settings->cpu_opt & CPU_SSE2)
+		RFX_INIT_SIMD(context);
 
 	return context;
 }

libfreerdp-gdi/gdi.c

@@ -1054,7 +1054,7 @@ int gdi_init(freerdp* instance, uint32 flags, uint8* buffer)
 
 	gdi_register_graphics(instance->context->graphics);
 
-	gdi->rfx_context = rfx_context_new();
+	gdi->rfx_context = rfx_context_new(instance->settings);
 	gdi->nsc_context = nsc_context_new();
 
 	return 0;

server/X11/xf_peer.c

@@ -121,7 +121,7 @@ void xf_peer_context_size(freerdp_peer* client, uint32* size)
 void xf_peer_context_new(freerdp_peer* client, xfPeerContext* context)
 {
 	context->info = xf_info_init();
-	context->rfx_context = rfx_context_new();
+	context->rfx_context = rfx_context_new(client->settings);
 	context->rfx_context->mode = RLGR3;
 	context->rfx_context->width = client->settings->width;
 	context->rfx_context->height = client->settings->height;

server/test/tfreerdp.c

@@ -64,7 +64,7 @@ void test_peer_context_size(freerdp_peer* client, uint32* size)
 
 void test_peer_context_new(freerdp_peer* client, testPeerContext* context)
 {
-	context->rfx_context = rfx_context_new();
+	context->rfx_context = rfx_context_new(client->settings);
 	context->rfx_context->mode = RLGR3;
 	context->rfx_context->width = client->settings->width;
 	context->rfx_context->height = client->settings->height;