From fcf83dbf9effab3bd98bad2b83b2468b7eb05cfd Mon Sep 17 00:00:00 2001
From: Tim Foley <tfoley@nvidia.com>
Date: Fri, 9 Jun 2017 11:34:21 -0700
Subject: Initial import of code.

---
 .../dxsdk/HDRToneMappingCS11/ReduceToSingleCS.hlsl | 63 ++++++++++++++++++++++
 1 file changed, 63 insertions(+)
 create mode 100644 tests/hlsl/dxsdk/HDRToneMappingCS11/ReduceToSingleCS.hlsl

(limited to 'tests/hlsl/dxsdk/HDRToneMappingCS11/ReduceToSingleCS.hlsl')

diff --git a/tests/hlsl/dxsdk/HDRToneMappingCS11/ReduceToSingleCS.hlsl b/tests/hlsl/dxsdk/HDRToneMappingCS11/ReduceToSingleCS.hlsl
new file mode 100644
index 000000000..cf506283e
--- /dev/null
+++ b/tests/hlsl/dxsdk/HDRToneMappingCS11/ReduceToSingleCS.hlsl
@@ -0,0 +1,63 @@
+//TEST:COMPARE_HLSL: -target dxbc-assembly -profile cs_4_0 -entry CSMain
+//-----------------------------------------------------------------------------
+// File: ReduceToSingleCS.hlsl
+//
+// Desc: Reduce an input buffer by a factor of groupthreads
+// 
+// Copyright (c) Microsoft Corporation. All rights reserved.
+//-----------------------------------------------------------------------------
+
+StructuredBuffer<float> Input : register( t0 );
+RWStructuredBuffer<float> Result : register( u0 );
+
+cbuffer cbCS : register( b0 )
+{
+    uint4    g_param;   // g_param.x is the actual elements contained in Input
+                        // g_param.y is the x dimension of the Dispatch call
+};
+
+#define groupthreads 128
+groupshared float accum[groupthreads];
+
+[numthreads(groupthreads,1,1)]
+void CSMain( uint3 Gid : SV_GroupID, uint3 DTid : SV_DispatchThreadID, uint3 GTid : SV_GroupThreadID, uint GI : SV_GroupIndex )
+{
+    if ( DTid.x < g_param.x )
+        accum[GI] = Input[DTid.x];
+    else
+        accum[GI] = 0;
+
+    // Parallel reduction algorithm follows 
+    GroupMemoryBarrierWithGroupSync();
+    if ( GI < 64 )
+        accum[GI] += accum[64+GI];  
+
+    GroupMemoryBarrierWithGroupSync();
+    if ( GI < 32 )    
+        accum[GI] += accum[32+GI];
+
+    GroupMemoryBarrierWithGroupSync();
+    if ( GI < 16 )
+        accum[GI] += accum[16+GI];
+
+    GroupMemoryBarrierWithGroupSync();
+    if ( GI < 8 ) 
+        accum[GI] += accum[8+GI];
+
+    GroupMemoryBarrierWithGroupSync();
+    if ( GI < 4 )
+        accum[GI] += accum[4+GI];
+
+    GroupMemoryBarrierWithGroupSync();
+    if ( GI < 2 )
+        accum[GI] += accum[2+GI];
+
+    GroupMemoryBarrierWithGroupSync();
+    if ( GI < 1 )
+        accum[GI] += accum[1+GI];
+    
+    if ( GI == 0 )
+    {        
+        Result[Gid.x] = accum[0];
+    }
+}
-- 
cgit v1.2.3