Pass CountingPushData into CountingParameters

Author: nipunG314

CountingSort/app_resources/prefix_sum_shader.comp.hlsl

@@ -42,12 +42,18 @@ uint32_t3 nbl::hlsl::glsl::gl_WorkGroupSize()
 [numthreads(WorkgroupSize,1,1)]
 void main(uint32_t3 ID : SV_GroupThreadID, uint32_t3 GroupID : SV_GroupID)
 {
+    nbl::hlsl::sort::CountingParameters < uint32_t > params;
+    params.dataElementCount = pushData.dataElementCount;
+    params.elementsPerWT = pushData.elementsPerWT;
+    params.minimum = pushData.minimum;
+    params.maximum = pushData.maximum;
+
     nbl::hlsl::sort::counting <uint32_t, PtrAccessor, PtrAccessor, PtrAccessor> counter;
     PtrAccessor input_accessor = PtrAccessor::create(pushData.inputKeyAddress);
     PtrAccessor scratch_accessor = PtrAccessor::create(pushData.scratchAddress);
-    /*counter.histogram(
+    counter.histogram(
         input_accessor,
         scratch_accessor,
-        pushData
-    );*/
+        params
+    );
 }

CountingSort/app_resources/scatter_shader.comp.hlsl

@@ -67,14 +67,20 @@ uint32_t3 nbl::hlsl::glsl::gl_WorkGroupSize()
 [numthreads(WorkgroupSize, 1, 1)]
 void main(uint32_t3 ID : SV_GroupThreadID, uint32_t3 GroupID : SV_GroupID)
 {
+    nbl::hlsl::sort::CountingParameters < uint32_t > params;
+    params.dataElementCount = pushData.dataElementCount;
+    params.elementsPerWT = pushData.elementsPerWT;
+    params.minimum = pushData.minimum;
+    params.maximum = pushData.maximum;
+
     nbl::hlsl::sort::counting <uint32_t, DoublePtrAccessor, DoublePtrAccessor, PtrAccessor> counter;
     DoublePtrAccessor key_accessor = DoublePtrAccessor::create(pushData.inputKeyAddress, pushData.outputKeyAddress);
     DoublePtrAccessor value_accessor = DoublePtrAccessor::create(pushData.inputValueAddress, pushData.outputValueAddress);
     PtrAccessor scratch_accessor = PtrAccessor::create(pushData.scratchAddress);
-    /*counter.scatter(
+    counter.scatter(
         key_accessor,
         value_accessor,
         scratch_accessor,
-        pushData
-    );*/
+        params
+    );
 }

CountingSort/main.cpp

@@ -32,8 +32,8 @@ class CountingSortApp final : public application_templates::MonoDeviceApplicatio
 		auto limits = m_physicalDevice->getLimits();
 		const uint32_t WorkgroupSize = limits.maxComputeWorkGroupInvocations;
 		const uint32_t MaxBucketCount = (limits.maxComputeSharedMemorySize / sizeof(uint32_t)) / 2;
-		constexpr size_t element_count = 100000;
-		const size_t bucket_count = std::min((uint32_t)3000, MaxBucketCount);
+		constexpr uint32_t element_count = 100000;
+		const uint32_t bucket_count = std::min((uint32_t)3000, MaxBucketCount);
 		const uint32_t elements_per_thread = ceil((float)ceil((float)element_count / limits.computeUnits) / WorkgroupSize);
 
 		// this time we load a shader directly from a file
@@ -171,7 +171,6 @@ class CountingSortApp final : public application_templates::MonoDeviceApplicatio
 
 		// Generate random data
 		constexpr uint32_t minimum = 0;
-		constexpr uint32_t maximum = 10000;
 		const uint32_t range = bucket_count;
 		unsigned seed = std::chrono::system_clock::now().time_since_epoch().count();
 		std::mt19937 g(seed);
@@ -212,7 +211,7 @@ class CountingSortApp final : public application_templates::MonoDeviceApplicatio
 			.dataElementCount = element_count,
 			.elementsPerWT = elements_per_thread,
 			.minimum = minimum,
-			.maximum = maximum,
+			.maximum = minimum + bucket_count - 1,
 		};
 
 		smart_refctd_ptr<nbl::video::IGPUCommandBuffer> cmdBuf;