diff --git a/lib/clover_deriv_quda.cu b/lib/clover_deriv_quda.cu
index 34ef0b993b..24615f84ce 100644
--- a/lib/clover_deriv_quda.cu
+++ b/lib/clover_deriv_quda.cu
@@ -12,6 +12,7 @@ namespace quda {
     double coeff;
     int parity;
     unsigned int minThreads() const { return gauge.LocalVolumeCB(); }
+    unsigned int sharedBytesPerThread() const { return 4 * sizeof(int); } // for thread_array
 
   public:
     DerivativeClover(GaugeField &force, GaugeField &gauge, GaugeField &oprod, double coeff, int parity) :
diff --git a/lib/coarse_op.cuh b/lib/coarse_op.cuh
index 564d1ed55c..78d5cd716e 100644
--- a/lib/coarse_op.cuh
+++ b/lib/coarse_op.cuh
@@ -245,7 +245,7 @@ namespace quda {
 
     unsigned int sharedBytesPerBlock(const TuneParam &param) const override
     {
-      if (type == COMPUTE_VUV || type == COMPUTE_VLV)
+      if (arg.shared_atomic && (type == COMPUTE_VUV || type == COMPUTE_VLV))
         return 4*sizeof(storeType)*arg.max_color_height_per_block*arg.max_color_width_per_block*4*coarseSpin*coarseSpin;
       return TunableKernel3D::sharedBytesPerBlock(param);
     }
@@ -577,9 +577,7 @@ namespace quda {
       if (type == COMPUTE_VUV || type == COMPUTE_VLV || type == COMPUTE_CONVERT || type == COMPUTE_RESCALE) arg.dim_index = 4*(dir==QUDA_BACKWARDS ? 0 : 1) + dim;
       arg.kd_dagger = kd_dagger;
 
-      if (type == COMPUTE_VUV || type == COMPUTE_VLV) tp.shared_bytes -= sharedBytesPerBlock(tp); // shared memory is static so don't include it in launch
       Launch<location_template>(arg, tp, type, stream);
-      if (type == COMPUTE_VUV || type == COMPUTE_VLV) tp.shared_bytes += sharedBytesPerBlock(tp); // restore shared memory
     };
 
     /**
diff --git a/lib/dslash5_domain_wall.cu b/lib/dslash5_domain_wall.cu
index 4dea657a25..f16b6cd7ab 100644
--- a/lib/dslash5_domain_wall.cu
+++ b/lib/dslash5_domain_wall.cu
@@ -68,7 +68,9 @@ namespace quda
     int blockMin() const { return 4; }
     unsigned int sharedBytesPerThread() const
     {
-      if (mobius_m5::shared()) {
+      if (mobius_m5::shared()
+          && (type == Dslash5Type::M5_INV_DWF || type == Dslash5Type::M5_INV_MOBIUS
+              || type == Dslash5Type::M5_INV_ZMOBIUS)) {
         // spin components in shared depend on inversion algorithm
 	bool isInv = type == Dslash5Type::M5_INV_DWF || type == Dslash5Type::M5_INV_MOBIUS || type == Dslash5Type::M5_INV_ZMOBIUS;
         int nSpin = (!isInv || mobius_m5::var_inverse()) ? mobius_m5::use_half_vector() ? in.Nspin() / 2 : in.Nspin() : in.Nspin();
@@ -81,7 +83,9 @@ namespace quda
     // overloaded to return max dynamic shared memory if doing shared-memory inverse
     unsigned int maxSharedBytesPerBlock() const
     {
-      if (mobius_m5::shared()) {
+      if (mobius_m5::shared()
+          && (type == Dslash5Type::M5_INV_DWF || type == Dslash5Type::M5_INV_MOBIUS
+              || type == Dslash5Type::M5_INV_ZMOBIUS)) {
         return maxDynamicSharedBytesPerBlock();
       } else {
         return TunableKernel3D::maxSharedBytesPerBlock();
@@ -104,7 +108,9 @@ namespace quda
       xpay(a == 0.0 ? false : true),
       type(type)
     {
-      if (mobius_m5::shared()) {
+      if (mobius_m5::shared()
+          && (type == Dslash5Type::M5_INV_DWF || type == Dslash5Type::M5_INV_MOBIUS
+              || type == Dslash5Type::M5_INV_ZMOBIUS)) {
         TunableKernel2D_base<false>::resizeStep(in.X(4)); // Ls must be contained in the block
       }
 
diff --git a/lib/gauge_ape.cu b/lib/gauge_ape.cu
index 248b7d1d6c..a2c8a92dd8 100644
--- a/lib/gauge_ape.cu
+++ b/lib/gauge_ape.cu
@@ -13,6 +13,7 @@ namespace quda {
     const GaugeField &in;
     const Float alpha;
     unsigned int minThreads() const { return in.LocalVolumeCB(); }
+    unsigned int sharedBytesPerThread() const { return 4 * sizeof(int); } // for thread_array
 
   public:
     // (2,3): 2 for parity in the y thread dim, 3 corresponds to mapping direction to the z thread dim
diff --git a/lib/gauge_stout.cu b/lib/gauge_stout.cu
index c7f256f2ee..48d7e638e8 100644
--- a/lib/gauge_stout.cu
+++ b/lib/gauge_stout.cu
@@ -18,8 +18,9 @@ namespace quda {
     unsigned int maxSharedBytesPerBlock() const { return maxDynamicSharedBytesPerBlock(); }
     unsigned int sharedBytesPerThread() const
     {
-      // use SharedMemoryCache if using over improvement for two link fields
-      return improved ? 2 * in.Ncolor() * in.Ncolor() * 2 * sizeof(typename mapper<Float>::type) : 0;
+      // use ThreadLocalCache if using over improvement for two link fields
+      return (improved ? 2 * in.Ncolor() * in.Ncolor() * 2 * sizeof(typename mapper<Float>::type) : 0)
+        + 4 * sizeof(int); // for thread_array
     }
 
   public:
diff --git a/lib/gauge_wilson_flow.cu b/lib/gauge_wilson_flow.cu
index 0c947e4e95..78456b665f 100644
--- a/lib/gauge_wilson_flow.cu
+++ b/lib/gauge_wilson_flow.cu
@@ -24,8 +24,9 @@ namespace quda {
 
     unsigned int sharedBytesPerThread() const
     {
-      // use SharedMemoryCache if using Symanzik improvement for two Link fields
-      return 4*sizeof(int) + (wflow_type == QUDA_GAUGE_SMEAR_SYMANZIK_FLOW ? 2 * in.Ncolor() * in.Ncolor() * 2 * sizeof(typename mapper<Float>::type) : 0);
+      // use ThreadLocalCache if using Symanzik improvement for two Link fields
+      return (wflow_type == QUDA_GAUGE_SMEAR_SYMANZIK_FLOW ? 2 * in.Ncolor() * in.Ncolor() * 2 * sizeof(typename mapper<Float>::type) : 0)
+        + 4 * sizeof(int); // for thread_array
     }
 
   public: