diff --git a/CUETools.Codecs.FLACCL/flac.cl b/CUETools.Codecs.FLACCL/flac.cl
index 63f2d44..1d8511d 100644
--- a/CUETools.Codecs.FLACCL/flac.cl
+++ b/CUETools.Codecs.FLACCL/flac.cl
@@ -474,13 +474,6 @@ void clEstimateResidual(
 
     barrier(CLK_LOCAL_MEM_FENCE);
 
-#ifdef AMD
-    float4 cptr0 = vload4(0, &fcoef[0]);
-    float4 cptr1 = vload4(1, &fcoef[0]);
-#if MAX_ORDER > 8
-    float4 cptr2 = vload4(2, &fcoef[0]);
-#endif
-#endif
     for (int pos = 0; pos < bs; pos += GROUP_SIZE)
     {
 	// fetch samples
@@ -491,11 +484,10 @@ void clEstimateResidual(
 
 	// compute residual
 	__local float* dptr = &data[tid + GROUP_SIZE - ro];
-#ifdef AMD
-	float4 sum = cptr0 * vload4(0, dptr)
-	    + cptr1 * vload4(1, dptr)
+	float4 sum = vload4(0, &fcoef[0]) * vload4(0, dptr)
+	    + vload4(1, &fcoef[0]) * vload4(1, dptr)
 #if MAX_ORDER > 8
-	    + cptr2 * vload4(2, dptr)
+	    + vload4(2, &fcoef[0]) * vload4(2, dptr)
   #if MAX_ORDER > 12
 	    + vload4(3, &fcoef[0]) * vload4(3, dptr)
     #if MAX_ORDER > 16
@@ -509,25 +501,33 @@ void clEstimateResidual(
 	    ;
 
 	int t = convert_int_rte(nextData + sum.x + sum.y + sum.z + sum.w);
-#else
-	float sum = 
-	    fcoef[0] * dptr[0] + fcoef[1] * dptr[1] + fcoef[2] * dptr[2] + fcoef[3] * dptr[3] + 
-	    fcoef[4] * dptr[4] + fcoef[5] * dptr[5] + fcoef[6] * dptr[6] + fcoef[7] * dptr[7] + 
-	    fcoef[8] * dptr[8] + fcoef[9] * dptr[9] + fcoef[10] * dptr[10] + fcoef[11] * dptr[11] ;
-	int t = convert_int_rte(nextData + sum);
-#endif
 	barrier(CLK_LOCAL_MEM_FENCE);
-	data[tid] = nextData;
 	// ensure we're within frame bounds
 	t = select(0, t, offs >= ro && offs < bs);
 	// overflow protection
 	t = iclamp(t, -0x7fffff, 0x7fffff);
 	// convert to unsigned
-	//if (offs < bs)
-	    atom_add(&psum[offs >> partOrder], (t << 1) ^ (t >> 31));
+#ifdef AMD
+	data[tid] = nextData;
+	atom_add(&psum[min(63,offs >> partOrder)], (t << 1) ^ (t >> 31));
+#else
+	data[tid] = (t << 1) ^ (t >> 31);
+	barrier(CLK_LOCAL_MEM_FENCE);
+	int ps = (1 << partOrder) - 1;
+	for (int l = 1 << (partOrder - 1); l > 0; l >>= 1)
+	{
+             if ((tid & ps) < l) 
+		data[tid] += data[tid + l];
+	    barrier(CLK_LOCAL_MEM_FENCE);
+	}
+	if ((tid & ps) == 0)
+		psum[min(63,offs >> partOrder)] += data[tid];
+	data[tid] = nextData;
+#endif
     }
 
     // calculate rice partition bit length for every (1 << partOrder) samples
+    barrier(CLK_LOCAL_MEM_FENCE);
     if (tid < 64)
     {
 	int k = iclamp(clz(1 << partOrder) - clz(psum[tid]), 0, 14); // 27 - clz(res) == clz(16) - clz(res) == log2(res / 16)
diff --git a/CUETools.Codecs.FLACCL/flaccpu.cl b/CUETools.Codecs.FLACCL/flaccpu.cl
index 0aa06a1..8b3f562 100644
--- a/CUETools.Codecs.FLACCL/flaccpu.cl
+++ b/CUETools.Codecs.FLACCL/flaccpu.cl
@@ -131,8 +131,9 @@ void clFindWastedBits(
     }
 }
 
-#define TEMPBLOCK 64
+#define TEMPBLOCK 128
 
+#if 0
 // get_num_groups(0) == number of tasks
 // get_num_groups(1) == number of windows
 __kernel __attribute__((reqd_work_group_size(1, 1, 1)))
@@ -173,6 +174,78 @@ void clComputeAutocor(
     for (int i = 0; i <= MAX_ORDER; ++i)
 	pout[i] = ac[i];
 }
+#else
+#define STORE_AC(ro, val) if (ro <= MAX_ORDER) pout[ro] = val;
+#define STORE_AC4(ro, val) STORE_AC(ro*4+0, val##ro.x) STORE_AC(ro*4+1, val##ro.y) STORE_AC(ro*4+2, val##ro.z) STORE_AC(ro*4+3, val##ro.w)
+
+// get_num_groups(0) == number of tasks
+// get_num_groups(1) == number of windows
+__kernel __attribute__((reqd_work_group_size(1, 1, 1)))
+void clComputeAutocor(
+    __global float *output,
+    __global const int *samples,
+    __global const float *window,
+    __global FLACCLSubframeTask *tasks,
+    const int taskCount // tasks per block
+)
+{
+    FLACCLSubframeData task = tasks[get_group_id(0) * taskCount].data;
+    int len = task.blocksize;
+    int windowOffs = get_group_id(1) * len;
+    float data[TEMPBLOCK + MAX_ORDER + 3];
+    double4 ac0 = 0.0, ac1 = 0.0, ac2 = 0.0, ac3 = 0.0, ac4 = 0.0, ac5 = 0.0, ac6 = 0.0, ac7 = 0.0, ac8 = 0.0;
+    
+    for (int pos = 0; pos < len; pos += TEMPBLOCK)
+    {
+	for (int tid = 0; tid < TEMPBLOCK + MAX_ORDER + 3; tid++)
+	    data[tid] = tid < len - pos ? samples[task.samplesOffs + pos + tid] * window[windowOffs + pos + tid] : 0.0f;
+
+	for (int j = 0; j < TEMPBLOCK;)
+	{
+    	    float4 temp0 = 0.0f, temp1 = 0.0f, temp2 = 0.0f, temp3 = 0.0f, temp4 = 0.0f, temp5 = 0.0f, temp6 = 0.0f, temp7 = 0.0f, temp8 = 0.0f;
+	    for (int k = 0; k < 32; k++)
+	    {
+		float d0 = data[j];
+		temp0 += d0 * vload4(0, &data[j]);
+		temp1 += d0 * vload4(1, &data[j]);
+#if MAX_ORDER >= 8
+		temp2 += d0 * vload4(2, &data[j]);
+#if MAX_ORDER >= 12
+		temp3 += d0 * vload4(3, &data[j]);
+#if MAX_ORDER >= 16
+		temp4 += d0 * vload4(4, &data[j]);
+		temp5 += d0 * vload4(5, &data[j]);
+		temp6 += d0 * vload4(6, &data[j]);
+		temp7 += d0 * vload4(7, &data[j]);
+		temp8 += d0 * vload4(8, &data[j]);
+#endif
+#endif
+#endif
+		j++;
+	    }
+	    ac0 += convert_double4(temp0);
+	    ac1 += convert_double4(temp1);
+    #if MAX_ORDER >= 8
+	    ac2 += convert_double4(temp2);
+    #if MAX_ORDER >= 12
+	    ac3 += convert_double4(temp3);
+    #if MAX_ORDER >= 16
+	    ac4 += convert_double4(temp4);
+	    ac5 += convert_double4(temp5);
+	    ac6 += convert_double4(temp6);
+	    ac7 += convert_double4(temp7);
+	    ac8 += convert_double4(temp8);
+    #endif
+    #endif
+    #endif
+	}
+    }
+    __global float * pout = &output[(get_group_id(0) * get_num_groups(1) + get_group_id(1)) * (MAX_ORDER + 1)];
+    STORE_AC4(0, ac) STORE_AC4(1, ac) STORE_AC4(2, ac) STORE_AC4(3, ac)
+    STORE_AC4(4, ac) STORE_AC4(5, ac) STORE_AC4(6, ac) STORE_AC4(7, ac)
+    STORE_AC4(8, ac)
+}
+#endif
 
 __kernel __attribute__((reqd_work_group_size(1, 1, 1)))
 void clComputeLPC(