Changed position of memory barrier in workgroup FFT

Fletterio · Fletterio · commit 5777e2abc9b2 · 2024-08-04T18:10:00.000-03:00
diff --git a/examples_tests b/examples_tests
@@ -1 +1 @@
-Subproject commit c3cb5351e28afae86e62c87dc1cc76f0af2f7fdd
+Subproject commit 28f3ea5f4e117c762aa7a0e686eb7dda18726852
diff --git a/include/nbl/builtin/hlsl/workgroup/fft.hlsl b/include/nbl/builtin/hlsl/workgroup/fft.hlsl
@@ -219,8 +219,9 @@ struct FFT<K, true, Scalar, device_capabilities>
         }
         accessor = offsetAccessor.accessor;
       
-        for (uint32_t stride = _NBL_HLSL_WORKGROUP_SIZE_ << 1; stride <= (K >> 1) * _NBL_HLSL_WORKGROUP_SIZE_; stride <<= 1)
+        for (uint32_t stride = 2 * _NBL_HLSL_WORKGROUP_SIZE_; stride < K * _NBL_HLSL_WORKGROUP_SIZE_; stride <<= 1)
         {
+            accessor.memoryBarrier(); // no execution barrier just making sure writes propagate to accessor
             //[unroll(K/2)]
             for (uint32_t virtualThreadID = SubgroupContiguousIndex(); virtualThreadID < (K >> 1) * _NBL_HLSL_WORKGROUP_SIZE_; virtualThreadID += _NBL_HLSL_WORKGROUP_SIZE_)
             {
@@ -243,7 +244,7 @@ struct FFT<K, true, Scalar, device_capabilities>
                 accessor.set(loIx, lo);
                 accessor.set(hiIx, hi);
             }
-            accessor.memoryBarrier(); // no execution barrier just making sure writes propagate to accessor
+           
         }
         
     }

Original file line number	Diff line number	Diff line change
`@@ -219,8 +219,9 @@ struct FFT<K, true, Scalar, device_capabilities>`
`219`	`219`	`}`
`220`	`220`	`accessor = offsetAccessor.accessor;`
`221`	`221`
`222`		`- for (uint32_t stride = _NBL_HLSL_WORKGROUP_SIZE_ << 1; stride <= (K >> 1) * _NBL_HLSL_WORKGROUP_SIZE_; stride <<= 1)`
	`222`	`+ for (uint32_t stride = 2 * _NBL_HLSL_WORKGROUP_SIZE_; stride < K * _NBL_HLSL_WORKGROUP_SIZE_; stride <<= 1)`
`223`	`223`	`{`
	`224`	`+ accessor.memoryBarrier(); // no execution barrier just making sure writes propagate to accessor`
`224`	`225`	`//[unroll(K/2)]`
`225`	`226`	`for (uint32_t virtualThreadID = SubgroupContiguousIndex(); virtualThreadID < (K >> 1) * _NBL_HLSL_WORKGROUP_SIZE_; virtualThreadID += _NBL_HLSL_WORKGROUP_SIZE_)`
`226`	`227`	`{`
`@@ -243,7 +244,7 @@ struct FFT<K, true, Scalar, device_capabilities>`
`243`	`244`	`accessor.set(loIx, lo);`
`244`	`245`	`accessor.set(hiIx, hi);`
`245`	`246`	`}`
`246`		`- accessor.memoryBarrier(); // no execution barrier just making sure writes propagate to accessor`
	`247`	`+`
`247`	`248`	`}`
`248`	`249`
`249`	`250`	`}`