Sound: added optimized x86_64 mixing for mingw and unix builds

MSVC2017: enabled multi-core compilation (-faster builds) Misc: slight speedup of main menu rendering
ec- · Aug 19, 2023 · 74648c0 · 74648c0
1 parent 1b06509
commit 74648c0
Show file tree

Hide file tree

Showing 13 changed files with 191 additions and 3 deletions.
diff --git a/Makefile b/Makefile
@@ -1105,6 +1105,11 @@ ifndef MINGW
 endif
 endif
 
+ifeq ($(ARCH),x86_64)
+  Q3OBJ += \
+    $(B)/client/snd_mix_x86_64.o
+endif
+
 ifeq ($(HAVE_VM_COMPILED),true)
   ifeq ($(ARCH),x86)
     Q3OBJ += $(B)/client/vm_x86.o

diff --git a/code/asm/snd_mix_x86_64.s b/code/asm/snd_mix_x86_64.s
@@ -0,0 +1,126 @@
+#if defined(MACOS_X)
+#undef ELF
+#endif
+
+#ifdef __ELF__
+.section .note.GNU-stack,"",@progbits
+#endif
+
+#ifdef ELF
+#define C(label) label
+#else
+#define C(label) _##label
+#endif
+
+.intel_syntax noprefix
+
+.text
+
+.globl C(S_WriteLinearBlastStereo16_SSE_x64)
+C(S_WriteLinearBlastStereo16_SSE_x64):
+	push rbx
+#ifdef _WIN32
+	// RCX RDX R8 R9
+	push rsi
+	push rdi
+	mov rsi, rcx
+	mov rdi, rdx
+	mov ebx, r8d
+#else
+	// RDI RSI RDX RCX R8 R9
+	xchg rsi, rdi
+	mov ebx, edx
+#endif
+	test ebx,ebx
+	jz	LExit
+	mov ecx,esi
+	and ecx,63
+	jz LMain
+	and ecx,3
+	jnz LTail
+	shr ecx,2
+	not ecx
+	add ecx,17
+LClamp1:
+	mov eax,[rsi]
+	sar eax,8
+	cmp eax,32767
+	jg	LClampHigh1
+	cmp eax,-32768
+	jnl LClampDone1
+	mov eax,-32768
+	jmp LClampDone1
+LClampHigh1:
+	mov eax,32767
+LClampDone1:
+	mov [rdi],ax
+	add rsi,4
+	add rdi,2
+	dec ebx
+	jz	LExit
+	dec ecx
+	jnz	LClamp1
+LMain:
+	mov ecx,ebx
+	shr ecx,4
+	jz  LTail
+	and ebx,15
+LAgain:
+	movq mm0, qword ptr [rsi+ 0]
+	movq mm1, qword ptr [rsi+ 8]
+	movq mm2, qword ptr [rsi+16]
+	movq mm3, qword ptr [rsi+24]
+	movq mm4, qword ptr [rsi+32]
+	movq mm5, qword ptr [rsi+40]
+	movq mm6, qword ptr [rsi+48]
+	movq mm7, qword ptr [rsi+56]
+	psrad mm0,8
+	psrad mm1,8
+	psrad mm2,8
+	psrad mm3,8
+	psrad mm4,8
+	psrad mm5,8
+	psrad mm6,8
+	psrad mm7,8
+	packssdw mm0, mm1
+	packssdw mm2, mm3
+	packssdw mm4, mm5
+	packssdw mm6, mm7
+	movntq qword ptr [rdi+ 0], mm0
+	movntq qword ptr [rdi+ 8], mm2
+	movntq qword ptr [rdi+16], mm4
+	movntq qword ptr [rdi+24], mm6
+	add rsi, 64
+	add rdi, 32
+	dec ecx
+	jnz LAgain
+LTail:
+	test ebx, ebx
+	jz	LEnd
+LClamp2:
+	mov eax,[rsi]
+	sar eax,8
+	cmp eax,32767
+	jg	LClampHigh2
+	cmp eax,-32768
+	jnl LClampDone2
+	mov eax,-32768
+	jmp LClampDone2
+LClampHigh2:
+	mov eax,32767
+LClampDone2:
+	mov [rdi],ax
+	add rsi,4
+	add rdi,2
+	dec ebx
+	jnz	LClamp2
+LEnd:
+	sfence
+	emms
+LExit:
+#ifdef _WIN32
+	pop rdi
+	pop rsi
+#endif
+	pop rbx
+	ret
diff --git a/code/client/cl_scrn.c b/code/client/cl_scrn.c
@@ -530,8 +530,11 @@ static void SCR_DrawScreenField( stereoFrame_t stereoFrame ) {
 	// unless they are displaying game renderings
 	if ( uiFullscreen || cls.state < CA_LOADING ) {
 		if ( cls.glconfig.vidWidth * 480 > cls.glconfig.vidHeight * 640 ) {
+			// draw vertical bars on sides for legacy mods
+			const int w = (cls.glconfig.vidWidth - ((cls.glconfig.vidHeight * 640) / 480)) /2;
 			re.SetColor( g_color_table[ ColorIndex( COLOR_BLACK ) ] );
-			re.DrawStretchPic( 0, 0, cls.glconfig.vidWidth, cls.glconfig.vidHeight, 0, 0, 0, 0, cls.whiteShader );
+			re.DrawStretchPic( 0, 0, w, cls.glconfig.vidHeight, 0, 0, 0, 0, cls.whiteShader );
+			re.DrawStretchPic( cls.glconfig.vidWidth - w, 0, w, cls.glconfig.vidHeight, 0, 0, 0, 0, cls.whiteShader );
 			re.SetColor( NULL );
 		}
 	}

diff --git a/code/client/snd_mix.c b/code/client/snd_mix.c
@@ -265,7 +265,7 @@ __asm {
 
 #endif // id386
 
-#if idx64 && defined (_MSC_VER) && defined(USE_WIN32_ASM)
+#if idx64 && (!defined (_MSC_VER) || defined(USE_WIN32_ASM))
 void S_WriteLinearBlastStereo16_SSE_x64( int*, short*, int );
 #endif
 
@@ -299,7 +299,7 @@ void S_TransferStereo16( unsigned long *pbuf, int endtime )
 			S_WriteLinearBlastStereo16_MMX();
 		else
 #endif
-#if idx64 && defined (_MSC_VER) && defined (USE_WIN32_ASM)
+#if idx64 && (!defined (_MSC_VER) || defined (USE_WIN32_ASM))
 		S_WriteLinearBlastStereo16_SSE_x64( snd_p, snd_out, snd_linear_count );
 #else
 		S_WriteLinearBlastStereo16();

diff --git a/code/win32/msvc2017/botlib.vcxproj b/code/win32/msvc2017/botlib.vcxproj
@@ -106,20 +106,23 @@
       <WarningLevel>Level4</WarningLevel>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <ClCompile>
       <PreprocessorDefinitions>WIN32;NDEBUG;_LIB;BOTLIB;_CRT_SECURE_NO_DEPRECATE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <WarningLevel>Level4</WarningLevel>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">
     <ClCompile>
       <PreprocessorDefinitions>WIN32;NDEBUG;_LIB;BOTLIB;_CRT_SECURE_NO_DEPRECATE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <WarningLevel>Level4</WarningLevel>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
@@ -132,6 +135,7 @@
       <WarningLevel>Level3</WarningLevel>
       <FloatingPointExceptions>true</FloatingPointExceptions>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
@@ -143,6 +147,7 @@
       <BrowseInformation>true</BrowseInformation>
       <WarningLevel>Level3</WarningLevel>
       <FloatingPointExceptions>true</FloatingPointExceptions>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|ARM64'">
@@ -154,6 +159,7 @@
       <BrowseInformation>true</BrowseInformation>
       <WarningLevel>Level3</WarningLevel>
       <FloatingPointExceptions>true</FloatingPointExceptions>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemGroup>

diff --git a/code/win32/msvc2017/libjpeg.vcxproj b/code/win32/msvc2017/libjpeg.vcxproj
@@ -109,6 +109,7 @@
       <WarningLevel>Level3</WarningLevel>
       <FloatingPointExceptions>true</FloatingPointExceptions>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
@@ -119,6 +120,7 @@
       <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
       <WarningLevel>Level3</WarningLevel>
       <FloatingPointExceptions>true</FloatingPointExceptions>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|ARM64'">
@@ -129,6 +131,7 @@
       <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
       <WarningLevel>Level3</WarningLevel>
       <FloatingPointExceptions>true</FloatingPointExceptions>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
@@ -137,13 +140,15 @@
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <WarningLevel>Level3</WarningLevel>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <ClCompile>
       <PreprocessorDefinitions>WIN32;NDEBUG;_LIB;_CRT_SECURE_NO_DEPRECATE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <WarningLevel>Level3</WarningLevel>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">
@@ -152,6 +157,7 @@
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <WarningLevel>Level3</WarningLevel>
       <CompileAs>CompileAsC</CompileAs>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemGroup>

diff --git a/code/win32/msvc2017/libogg.vcxproj b/code/win32/msvc2017/libogg.vcxproj
@@ -114,6 +114,7 @@
       <PreprocessorDefinitions>WIN32;NDEBUG;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
     <Link>
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
@@ -128,6 +129,7 @@
       <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
       <FloatingPointExceptions>true</FloatingPointExceptions>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
@@ -137,13 +139,15 @@
       <PreprocessorDefinitions>WIN32;_DEBUG;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
       <FloatingPointExceptions>true</FloatingPointExceptions>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <ClCompile>
       <WarningLevel>Level3</WarningLevel>
       <PreprocessorDefinitions>WIN32;NDEBUG;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
     <Link>
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
@@ -155,6 +159,7 @@
       <PreprocessorDefinitions>WIN32;NDEBUG;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <WarningLevel>Level3</WarningLevel>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
     <Link>
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
@@ -169,6 +174,7 @@
       <BasicRuntimeChecks>EnableFastChecks</BasicRuntimeChecks>
       <FloatingPointExceptions>true</FloatingPointExceptions>
       <Optimization>Disabled</Optimization>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemGroup>

diff --git a/code/win32/msvc2017/libvorbis.vcxproj b/code/win32/msvc2017/libvorbis.vcxproj
@@ -114,6 +114,7 @@
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <WarningLevel>Level1</WarningLevel>
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
     <Link>
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
@@ -129,6 +130,7 @@
       <EnableEnhancedInstructionSet>NoExtensions</EnableEnhancedInstructionSet>
       <FloatingPointExceptions>true</FloatingPointExceptions>
       <BasicRuntimeChecks>EnableFastChecks</BasicRuntimeChecks>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
@@ -139,6 +141,7 @@
       <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
       <FloatingPointExceptions>true</FloatingPointExceptions>
       <BasicRuntimeChecks>EnableFastChecks</BasicRuntimeChecks>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
@@ -148,6 +151,7 @@
       <PreprocessorDefinitions>_CRT_SECURE_NO_WARNINGS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <WarningLevel>Level1</WarningLevel>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
     <Link>
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
@@ -162,13 +166,15 @@
       <FloatingPointExceptions>true</FloatingPointExceptions>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|ARM64'">
     <ClCompile>
       <PreprocessorDefinitions>_CRT_SECURE_NO_WARNINGS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
       <WarningLevel>Level1</WarningLevel>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
     </ClCompile>
     <Link>
       <EnableCOMDATFolding>true</EnableCOMDATFolding>