shithub: dumb

Download patch

ref: b5f940169c2968b54e8d48cf880f5999c5932922
parent: cf2577d4bc284fb5c5b71377413a47c72a1362dc
author: Chris Moeller <[email protected]>
date: Tue Jun 11 23:45:06 EDT 2013 

Re-enabled SSE support and rewrote SSE filter function


--- a/dumb/src/it/itrender.c
+++ b/dumb/src/it/itrender.c
@@ -747,19 +747,22 @@
 			dumb_record_click(cr, pos, startstep);

 		}

 

-		data = _mm_cvtsi32_ss( _mm_setzero_ps(), prevsample );

-		data = _mm_cvtsi32_ss( _mm_shuffle_ps( data, data, _MM_SHUFFLE(0, 0, 0, 0) ), currsample );

+		temp1 = _mm_setzero_ps();

+		data = _mm_cvtsi32_ss( temp1, currsample );

+		temp2 = _mm_cvtsi32_ss( temp1, prevsample );

 		impulse = _mm_loadu_ps( (const float *) &imp );

-		temp1 = _mm_shuffle_ps( data, data, _MM_SHUFFLE(0, 1, 0, 0) );

+		data = _mm_shuffle_ps( data, temp2, _MM_SHUFFLE(1, 0, 0, 1) );

 

 		for (i = 0; i < datasize; i += step) {

-			data = _mm_cvtsi32_ss( temp1, src [i] );

-			temp1 = _mm_mul_ps( data, impulse );

-			temp2 = _mm_movehl_ps( temp1, temp1 );

+			temp1 = _mm_cvtsi32_ss( data, src [i] );

+			temp1 = _mm_mul_ps( temp1, impulse );

+			temp2 = _mm_movehl_ps( temp2, temp1 );

 			temp1 = _mm_add_ps( temp1, temp2 );

-			temp2 = _mm_shuffle_ps( temp1, temp1, _MM_SHUFFLE(0, 0, 0, 1) );

+			temp2 = temp1;

+			temp2 = _mm_shuffle_ps( temp2, temp1, _MM_SHUFFLE(0, 0, 0, 1) );

 			temp1 = _mm_add_ps( temp1, temp2 );

-			temp1 = _mm_shuffle_ps( temp1, data, _MM_SHUFFLE(0, 1, 0, 0) );

+			temp1 = _mm_shuffle_ps( temp1, data, _MM_SHUFFLE(2, 1, 0, 0) );

+			data = temp1;

 			dst [i] += _mm_cvtss_si32( temp1 );

 		}

 

--- a/dumb/vc6/dumb/dumb.vcxproj
+++ b/dumb/vc6/dumb/dumb.vcxproj
@@ -55,7 +55,7 @@
     <ClCompile>

       <Optimization>Disabled</Optimization>

       <AdditionalIncludeDirectories>../../include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>

-      <PreprocessorDefinitions>_DEBUG;WIN32;_LIB;DUMB_DECLARE_DEPRECATED;DEBUGMODE=1;%(PreprocessorDefinitions)</PreprocessorDefinitions>

+      <PreprocessorDefinitions>_USE_SSE;_DEBUG;WIN32;_LIB;DUMB_DECLARE_DEPRECATED;DEBUGMODE=1;%(PreprocessorDefinitions)</PreprocessorDefinitions>

       <BasicRuntimeChecks>EnableFastChecks</BasicRuntimeChecks>

       <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>

       <WarningLevel>Level3</WarningLevel>

@@ -76,7 +76,7 @@
       <Optimization>MaxSpeed</Optimization>

       <InlineFunctionExpansion>AnySuitable</InlineFunctionExpansion>

       <AdditionalIncludeDirectories>../../include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>

-      <PreprocessorDefinitions>NDEBUG;WIN32;_LIB;DUMB_DECLARE_DEPRECATED;%(PreprocessorDefinitions)</PreprocessorDefinitions>

+      <PreprocessorDefinitions>_USE_SSE;NDEBUG;WIN32;_LIB;DUMB_DECLARE_DEPRECATED;%(PreprocessorDefinitions)</PreprocessorDefinitions>

       <StringPooling>true</StringPooling>

       <RuntimeLibrary>MultiThreaded</RuntimeLibrary>

       <FunctionLevelLinking>true</FunctionLevelLinking>