GPU buffer oriented: append global rays in parallel instead of only o…

…n thread 0 of the group; 619 -> 638 Mray/s
aras-p · Apr 25, 2018 · 88e3e9f · 88e3e9f
1 parent 30a62d2
commit 88e3e9f
Show file tree

Hide file tree

Showing 3 changed files with 29 additions and 14 deletions.
diff --git a/Cpp/Windows/ComputeShader.hlsl b/Cpp/Windows/ComputeShader.hlsl
@@ -233,6 +233,7 @@ RWStructuredBuffer<RayData> g_RayBufferDst : register(u2);
 // a CS group execution is appended into the global one.
 
 groupshared uint s_GroupRayCounter;
+groupshared uint s_GroupRayGlobalStart;
 #define kMaxGroupRays 768
 groupshared RayData s_GroupRays[kMaxGroupRays];
 
@@ -243,15 +244,19 @@ void PushRayData(RayData rd)
     s_GroupRays[index] = rd;
 }
 
-void PushGlobalRayData()
+void GetGlobalRayDataOffset(uint rayCount)
 {
-    // append new rays into global buffer
-    uint rayCount = min(s_GroupRayCounter, kMaxGroupRays);
-    uint rayBufferStart;
-    g_OutCounts.InterlockedAdd(4, rayCount, rayBufferStart);
-    for (uint ir = 0; ir < rayCount; ++ir)
+    g_OutCounts.InterlockedAdd(4, rayCount, s_GroupRayGlobalStart);
+}
+
+void PushGlobalRayData(uint threadID, uint rayCount, uint groupSize)
+{
+    uint myRayCount = (rayCount + groupSize - 1) / groupSize;
+    uint myRayStart = threadID * myRayCount;
+    for (uint ir = myRayStart; ir < myRayStart + myRayCount; ++ir)
     {
-        g_RayBufferDst[rayBufferStart + ir] = s_GroupRays[ir];
+        if (ir < rayCount)
+            g_RayBufferDst[s_GroupRayGlobalStart + ir] = s_GroupRays[ir];
     }
 }
 

diff --git a/Cpp/Windows/ComputeShaderBounce.hlsl b/Cpp/Windows/ComputeShaderBounce.hlsl
@@ -37,7 +37,8 @@ RWStructuredBuffer<SplatData> g_SplatBufferDst : register(u3);
 [numthreads(kCSRayBatchSize, 1, 1)]
 void main(uint3 gid : SV_DispatchThreadID, uint3 tid : SV_GroupThreadID)
 {
-    if (tid.x == 0)
+    uint threadID = tid.x;
+    if (threadID == 0)
     {
         s_GroupRayCounter = 0;
         s_GroupSplatCounter = 0;
@@ -91,12 +92,13 @@ void main(uint3 gid : SV_DispatchThreadID, uint3 tid : SV_GroupThreadID)
     //if (s_GroupRayCounter > kMaxGroupRays)
     //    PushSplat(float3(0, 2, 0), pixelIndex);
 
-    if (tid.x == 0)
+    uint rayCount = min(s_GroupRayCounter, kMaxGroupRays);
+    if (threadID == 0)
     {
         // total ray counts (for perf indicator)
         g_OutCounts.InterlockedAdd(0, kCSRayBatchSize);
 
-        PushGlobalRayData();
+        GetGlobalRayDataOffset(rayCount);
 
         // append new splats into global buffer
         uint splatBufferStart;
@@ -106,4 +108,6 @@ void main(uint3 gid : SV_DispatchThreadID, uint3 tid : SV_GroupThreadID)
             g_SplatBufferDst[splatBufferStart + is] = s_GroupSplats[is];
         }
     }
+    GroupMemoryBarrierWithGroupSync();
+    PushGlobalRayData(threadID, rayCount, kCSRayBatchSize);
 }
diff --git a/Cpp/Windows/ComputeShaderCameraRays.hlsl b/Cpp/Windows/ComputeShaderCameraRays.hlsl
@@ -1,9 +1,12 @@
 #include "ComputeShader.hlsl"
 
+groupshared uint s_RayBufferStart;
+
 [numthreads(kCSGroupSizeX, kCSGroupSizeY, 1)]
 void main(uint3 gid : SV_DispatchThreadID, uint3 tid : SV_GroupThreadID)
 {
-    if (tid.x == 0)
+    uint threadID = tid.x + tid.y * kCSGroupSizeX;
+    if (threadID == 0)
     {
         s_GroupRayCounter = 0;
     }
@@ -44,10 +47,13 @@ void main(uint3 gid : SV_DispatchThreadID, uint3 tid : SV_GroupThreadID)
     //if (s_GroupRayCounter > kMaxGroupRays)
     //    dstImage[gid.xy] += float4(2,0,0,0);
 
-    if (tid.x == 0 && tid.y == 0)
+    uint rayCount = min(s_GroupRayCounter, kMaxGroupRays);
+    if (threadID == 0)
     {
         g_OutCounts.InterlockedAdd(0, DO_SAMPLES_PER_PIXEL * kCSGroupSizeX * kCSGroupSizeY);
-
-        PushGlobalRayData();
+        GetGlobalRayDataOffset(rayCount);
     }
+    GroupMemoryBarrierWithGroupSync();
+
+    PushGlobalRayData(threadID, rayCount, kCSGroupSizeX*kCSGroupSizeY);
 }