Merge branch 'denormals-are-a-fuck' into 'master'

Ignore float denormals on ia32/64 for voice and sound

Closes ring-racers#283

See merge request kart-krew-dev/ring-racers-internal!2917

src/d_clisrv.c

@@ -7379,8 +7379,10 @@ void NetKeepAlive(void)
 	FileSendTicker();
 
 	// Update voice whenever possible.
+	{
 		NetVoiceUpdate();
 	}
+}
 
 // If a tree falls in the forest but nobody is around to hear it, does it make a tic?
 #define DEDICATEDIDLETIME (10*TICRATE)
@@ -7580,12 +7582,16 @@ void NetVoiceUpdate(void)
 	UINT8 *encoded = NULL;
 	float *subframe_buffer = NULL;
 	float *denoise_buffer = NULL;
+	ps_voiceupdatetime = I_GetPreciseTime();
 
 	if (dedicated)
 	{
+		ps_voiceupdatetime = I_GetPreciseTime() - ps_voiceupdatetime;
 		return;
 	}
 
+	floatdenormalstate_t dnzstate = M_EnterFloatDenormalToZero();
+
 	UINT32 bytes_dequed = 0;
 
 	bytes_dequed = S_SoundInputDequeueSamples((void*)(g_local_voice_buffer + g_local_voice_buffer_len), SRB2_VOICE_MAX_DEQUEUE_BYTES - (g_local_voice_buffer_len * sizeof(float)));
@@ -7724,9 +7730,12 @@ void NetVoiceUpdate(void)
 		g_local_voice_buffer_len -= buffer_offset;
 	}
 
+	M_ExitFloatDenormalToZero(dnzstate);
+
 	if (denoise_buffer) Z_Free(denoise_buffer);
 	if (subframe_buffer) Z_Free(subframe_buffer);
 	if (encoded) Z_Free(encoded);
+	ps_voiceupdatetime = I_GetPreciseTime() - ps_voiceupdatetime;
 	return;
 }
 

src/m_misc.cpp

@@ -32,6 +32,11 @@
 // Extended map support.
 #include <ctype.h>
 
+#if defined(__x86_64__) || defined(_M_X64) || defined(i386) || defined(__i386__) || defined(__i386) || defined(_M_IX86)
+#include <immintrin.h>
+#define NEED_INTEL_DENORMAL_BIT 1
+#endif
+
 #include "doomdef.h"
 #include "g_game.h"
 #include "m_misc.h"
@@ -2805,3 +2810,44 @@ const char * M_Ftrim (double f)
 		return &dig[1];/* skip the 0 */
 	}
 }
+
+/** Enable floating point denormal-to-zero section, if necessary */
+floatdenormalstate_t M_EnterFloatDenormalToZero(void)
+{
+#ifdef NEED_INTEL_DENORMAL_BIT
+	floatdenormalstate_t state = 0;
+	state |= _MM_GET_FLUSH_ZERO_MODE() == _MM_FLUSH_ZERO_ON ? 1 : 0;
+	state |= _MM_GET_DENORMALS_ZERO_MODE() == _MM_DENORMALS_ZERO_ON ? 2 : 0;
+
+	if ((state & 1) == 0)
+	{
+		_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
+	}
+	if ((state & 2) == 0)
+	{
+		_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+	}
+	return state;
+#else
+	return 0;
+#endif
+}
+
+/** Exit floating point denormal-to-zero section, if necessary, restoring previous state */
+void M_ExitFloatDenormalToZero(floatdenormalstate_t previous)
+{
+#ifdef NEED_INTEL_DENORMAL_BIT
+	if ((previous & 1) == 0)
+	{
+		_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_OFF);
+	}
+	if ((previous & 2) == 0)
+	{
+		_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_OFF);
+	}
+	return;
+#else
+	(void)previous;
+	return;
+#endif
+}

src/m_misc.h

@@ -167,6 +167,13 @@ FUNCMATH UINT8 M_CountBits(UINT32 num, UINT8 size);
 
 extern char configfile[MAX_WADPATH];
 
+typedef INT32 floatdenormalstate_t;
+
+/** Enable floating point denormal-to-zero section, if necessary */
+floatdenormalstate_t M_EnterFloatDenormalToZero(void);
+/** Exit floating point denormal-to-zero section, if necessary, restoring previous state */
+void M_ExitFloatDenormalToZero(floatdenormalstate_t previous);
+
 #ifdef __cplusplus
 } // extern "C"
 #endif

src/m_perfstats.c

@@ -61,6 +61,8 @@ int ps_checkposition_calls = 0;
 precise_t ps_lua_thinkframe_time = 0;
 int ps_lua_mobjhooks = 0;
 
+precise_t ps_voiceupdatetime = 0;
+
 // dynamically allocated resizeable array for thinkframe hook stats
 ps_hookinfo_t *thinkframe_hooks = NULL;
 int thinkframe_hooks_length = 0;
@@ -208,6 +210,7 @@ static void M_DrawRenderStats(void)
 
 	perfstatrow_t tictime_row[] = {
 		{"logic  ", "Game logic:    ", &ps_tictime},
+		{"voice  ", "Voice update:  ", &ps_voiceupdatetime},
 		{0}
 	};
 

src/m_perfstats.h

@@ -45,6 +45,8 @@ extern int       ps_checkposition_calls;
 extern precise_t ps_lua_thinkframe_time;
 extern int       ps_lua_mobjhooks;
 
+extern precise_t ps_voiceupdatetime;
+
 struct ps_hookinfo_t
 {
 	precise_t time_taken;

src/p_test.cpp

@@ -30,10 +30,10 @@ void P_TestLine(line_t* ld)
 	g_lines.emplace_back(ld);
 }
 
+extern "C" consvar_t cv_showgremlins;
+
 line_t* P_SweepTestLines(fixed_t ax, fixed_t ay, fixed_t bx, fixed_t by, fixed_t r, vector2_t* return_normal)
 {
-	extern consvar_t cv_showgremlins;
-
 	using namespace srb2::math;
 	using namespace srb2::sweep;
 

src/sdl/new_sound.cpp

@@ -31,6 +31,7 @@
 
 #include "../doomdef.h"
 #include "../i_sound.h"
+#include "../m_misc.h"
 #include "../s_sound.h"
 #include "../sounds.h"
 #include "../w_wad.h"
@@ -258,6 +259,8 @@ void audio_callback(void* userdata, Uint8* buffer, int len)
 	tracy::SetThreadName("SDL Audio Thread");
 	FrameMarkStart(kAudio);
 	ZoneScoped;
+	floatdenormalstate_t dtzstate = M_EnterFloatDenormalToZero();
+	auto dtzrestore = srb2::finally([dtzstate] { M_ExitFloatDenormalToZero(dtzstate); });
 
 	// The SDL Audio lock is implied to be held during callback.