stability issues

CUETools.Codecs.FLACCL/FLACCLWriter.cs

@@ -72,7 +72,7 @@ namespace CUETools.Codecs.FLACCL
 		public int GroupSize { get; set; }
 
 		[DefaultValue(32)]
-		[SRDescription(typeof(Properties.Resources), "DescriptionTashSize")]
+		[SRDescription(typeof(Properties.Resources), "DescriptionTaskSize")]
 		public int TaskSize { get; set; }
 
 		[SRDescription(typeof(Properties.Resources), "DescriptionDefines")]
@@ -845,18 +845,17 @@ namespace CUETools.Codecs.FLACCL
 
 			if (_settings.GPUOnly && _settings.DoRice)
 			{
+				int len = task.BestResidualTasks[index].size - task.BestResidualTasks[index].headerLen;
+				int pos = task.BestResidualTasks[index].encodingOffset;
 				if (task.BestResidualTasks[index].size != (int)sub.best.size)
 					throw new Exception("Encoding offset mismatch");
 				if (task.BestResidualTasks[index].headerLen != offs0 + 6)
 					throw new Exception("Encoding offset mismatch");
-				if (task.BestResidualTasks[index].encodingOffset != frame.writer.BitLength)
+				if (pos % 8 != frame.writer.BitLength % 8)
 					throw new Exception("Encoding offset mismatch");
-				int len = task.BestResidualTasks[index].size - task.BestResidualTasks[index].headerLen;
 				//Console.WriteLine("{0:x} => {1:x}", _totalSize + frame.writer.BitLength / 8, _totalSize + (frame.writer.BitLength + len) / 8);
 				// task.BestResidualTasks[index].headerLen
-				frame.writer.writeints(len, (byte*)task.clRiceOutputPtr);
+				frame.writer.writeints(len, pos, (byte*)task.clRiceOutputPtr);
-				if (task.BestResidualTasks[index].encodingOffset + len != frame.writer.BitLength)
-					throw new Exception("Encoding offset mismatch");
 			}
 			else
 			{
@@ -1212,8 +1211,15 @@ namespace CUETools.Codecs.FLACCL
 							}
 #endif
 
-							if (((csum << task.frame.subframes[ch].obits) >= 1UL << 32 && !_settings.DoRice) || !_settings.GPUOnly)
+							if (((csum << task.frame.subframes[ch].obits) >= 1UL << 32) || !_settings.GPUOnly)
 							{
+								if (_settings.GPUOnly && _settings.DoRice)
+#if DEBUG
+//									throw new Exception("DoRice failed");
+									break;
+#else
+									break;
+#endif
 								if (!unpacked) unpack_samples(task, task.frameSize); unpacked = true;
 								if ((csum << task.frame.subframes[ch].obits) >= 1UL << 32)
 									lpc.encode_residual_long(task.frame.subframes[ch].best.residual, task.frame.subframes[ch].samples, task.frame.blocksize, task.frame.subframes[ch].best.order, coefs, task.frame.subframes[ch].best.shift);
@@ -1255,7 +1261,7 @@ namespace CUETools.Codecs.FLACCL
 
 		unsafe void select_best_methods(FlacFrame frame, int channelsCount, int iFrame, FLACCLTask task)
 		{
-			if (channelsCount == 4 && channels == 2)
+			if (channelsCount == 4 && channels == 2 && frame.blocksize > 4)
 			{
 				if (task.BestResidualTasks[iFrame * 2].channel == 0 && task.BestResidualTasks[iFrame * 2 + 1].channel == 1)
 					frame.ch_mode = ChannelMode.LeftRight;
@@ -1321,7 +1327,7 @@ namespace CUETools.Codecs.FLACCL
 
 		unsafe void estimate_residual(FLACCLTask task, int channelsCount)
 		{
-			if (task.frameSize >= 4)
+			if (task.frameSize > 4)
 				task.EnqueueKernels();
 		}
 
@@ -1605,8 +1611,9 @@ namespace CUETools.Codecs.FLACCL
 				}
 				OCLMan.CreateDefaultContext(platformId, (DeviceType)_settings.DeviceType);
 
+				bool haveAtom = false;
 				if (OCLMan.Context.Devices[0].Extensions.Contains("cl_khr_local_int32_extended_atomics"))
-					_settings.Defines += "#define HAVE_ATOM\n";
+					haveAtom = true;					
 				else
 					_settings.GPUOnly = false;
 
@@ -1618,10 +1625,11 @@ namespace CUETools.Codecs.FLACCL
 					"#define FLACCL_VERSION \"" + vendor_string + "\"\n" +
 					(_settings.GPUOnly ? "#define DO_PARTITIONS\n" : "") +
 					(_settings.DoRice ? "#define DO_RICE\n" : "") +
+					(haveAtom ? "#define HAVE_ATOM\n" : "") +
 #if DEBUG
 					"#define DEBUG\n" +
 #endif
- (_settings.DeviceType == OpenCLDeviceType.CPU ? "#define FLACCL_CPU\n" : "") +
+					(_settings.DeviceType == OpenCLDeviceType.CPU ? "#define FLACCL_CPU\n" : "") +
 					_settings.Defines + "\n";
 
 				try

CUETools.Codecs.FLACCL/Properties/Resources.Designer.cs

@@ -1,7 +1,7 @@
 //------------------------------------------------------------------------------
 // <auto-generated>
 //     This code was generated by a tool.
-//     Runtime Version:2.0.50727.4200
+//     Runtime Version:2.0.50727.4206
 //
 //     Changes to this file may cause incorrect behavior and will be lost if
 //     the code is regenerated.
@@ -69,6 +69,33 @@ namespace CUETools.Codecs.FLACCL.Properties {
             }
         }
         
+        /// <summary>
+        ///   Looks up a localized string similar to Additional preprocessor definitions for OpenCL kernel.
+        /// </summary>
+        internal static string DescriptionDefines {
+            get {
+                return ResourceManager.GetString("DescriptionDefines", resourceCulture);
+            }
+        }
+        
+        /// <summary>
+        ///   Looks up a localized string similar to Use CPU or GPU device.
+        /// </summary>
+        internal static string DescriptionDeviceType {
+            get {
+                return ResourceManager.GetString("DescriptionDeviceType", resourceCulture);
+            }
+        }
+        
+        /// <summary>
+        ///   Looks up a localized string similar to Do final encoding stages on GPU (experimental).
+        /// </summary>
+        internal static string DescriptionDoRice {
+            get {
+                return ResourceManager.GetString("DescriptionDoRice", resourceCulture);
+            }
+        }
+        
         /// <summary>
         ///   Looks up a localized string similar to Use GPU on all stages.
         /// </summary>
@@ -78,6 +105,42 @@ namespace CUETools.Codecs.FLACCL.Properties {
             }
         }
         
+        /// <summary>
+        ///   Looks up a localized string similar to GPU thread block size (64, 128, 256).
+        /// </summary>
+        internal static string DescriptionGroupSize {
+            get {
+                return ResourceManager.GetString("DescriptionGroupSize", resourceCulture);
+            }
+        }
+        
+        /// <summary>
+        ///   Looks up a localized string similar to Device uses host memory (Don&apos;t use).
+        /// </summary>
+        internal static string DescriptionMappedMemory {
+            get {
+                return ResourceManager.GetString("DescriptionMappedMemory", resourceCulture);
+            }
+        }
+        
+        /// <summary>
+        ///   Looks up a localized string similar to OpenCL platform to use (ATI Stream, NVIDIA OpenCL, Intel OpenCL, etc).
+        /// </summary>
+        internal static string DescriptionPlatform {
+            get {
+                return ResourceManager.GetString("DescriptionPlatform", resourceCulture);
+            }
+        }
+        
+        /// <summary>
+        ///   Looks up a localized string similar to Number of frames processed simultaniously (32, 64).
+        /// </summary>
+        internal static string DescriptionTaskSize {
+            get {
+                return ResourceManager.GetString("DescriptionTaskSize", resourceCulture);
+            }
+        }
+        
         /// <summary>
         ///   Looks up a localized string similar to Calculate MD5 hash for audio stream.
         /// </summary>

CUETools.Codecs.FLACCL/Properties/Resources.resx

@@ -120,9 +120,30 @@
   <data name="DescriptionCPUThreads" xml:space="preserve">
     <value>Use additional CPU threads</value>
   </data>
+  <data name="DescriptionDefines" xml:space="preserve">
+    <value>Additional preprocessor definitions for OpenCL kernel</value>
+  </data>
+  <data name="DescriptionDeviceType" xml:space="preserve">
+    <value>Use CPU or GPU device</value>
+  </data>
+  <data name="DescriptionDoRice" xml:space="preserve">
+    <value>Do final encoding stages on GPU (experimental)</value>
+  </data>
   <data name="DescriptionGPUOnly" xml:space="preserve">
     <value>Use GPU on all stages</value>
   </data>
+  <data name="DescriptionGroupSize" xml:space="preserve">
+    <value>GPU thread block size (64, 128, 256)</value>
+  </data>
+  <data name="DescriptionMappedMemory" xml:space="preserve">
+    <value>Device uses host memory (Don't use)</value>
+  </data>
+  <data name="DescriptionPlatform" xml:space="preserve">
+    <value>OpenCL platform to use (ATI Stream, NVIDIA OpenCL, Intel OpenCL, etc)</value>
+  </data>
+  <data name="DescriptionTaskSize" xml:space="preserve">
+    <value>Number of frames processed simultaniously (32, 64)</value>
+  </data>
   <data name="DoMD5Description" xml:space="preserve">
     <value>Calculate MD5 hash for audio stream</value>
   </data>

CUETools.Codecs.FLACCL/flac.cl

@@ -1723,6 +1723,8 @@ void clCalcOutputOffsets(
 	{
 	    __global FLACCLSubframeTask* task = tasks + iFrame * channels + ch;
 	    offset += 8 + task->data.wbits;
+	    // Add 32 bits to separate frames if header is too small so they can intersect
+	    offset += 64;
 	    task->data.encodingOffset = offset + task->data.headerLen;
 	    offset += task->data.size;
 	}
@@ -1845,6 +1847,8 @@ void clRiceEncoding(
     if (tid < sizeof(task) / sizeof(int))
 	((__local int*)&task)[tid] = ((__global int*)(&tasks[get_group_id(0)]))[tid];
     barrier(CLK_LOCAL_MEM_FENCE);
+    if (task.type != Fixed && task.type != LPC)
+	return;
     if (tid == 0)
 	mypos[GROUP_SIZE] = 0;
     if (tid < WARP_SIZE)