Merge pull request #1380 from lattice/feature/sycl-merge

add ThreadLocalCache

include/kernels/block_transpose.cuh

@@ -47,6 +47,15 @@ namespace quda
     constexpr BlockTransposeKernel(const Arg &arg) : arg(arg) { }
     static constexpr const char *filename() { return KERNEL_FILE; }
 
+    struct CacheDims {
+      static constexpr dim3 dims(dim3 block)
+      {
+        block.x += 1;
+        block.z = 1;
+        return block;
+      }
+    };
+
     /**
       @brief Transpose between the two different orders of batched colorspinor fields:
         - B: nVec -> spatial/N -> spin/color -> N, where N is for that in floatN
@@ -60,7 +69,7 @@ namespace quda
       int parity = parity_color / Arg::nColor;
       using color_spinor_t = ColorSpinor<typename Arg::real, 1, Arg::nSpin>;
 
-      SharedMemoryCache<color_spinor_t> cache({target::block_dim().x + 1, target::block_dim().y, 1});
+      SharedMemoryCache<color_spinor_t, CacheDims> cache;
 
       int x_offset = target::block_dim().x * target::block_idx().x;
       int v_offset = target::block_dim().y * target::block_idx().y;

include/kernels/coarse_op_kernel.cuh

@@ -10,6 +10,7 @@
 #include <matrix_tile.cuh>
 #include <target_device.h>
 #include <kernel.h>
+#include <shared_memory_cache_helper.h>
 
 namespace quda {
 
@@ -1387,14 +1388,20 @@ namespace quda {
   };
 
   template <> struct storeCoarseSharedAtomic_impl<true> {
+    template <typename Arg>
+    using CacheT = complex<storeType>[Arg::max_color_height_per_block][Arg::max_color_width_per_block][4]
+                                     [Arg::coarseSpin][Arg::coarseSpin];
+    template <typename Arg> using Cache = SharedMemoryCache<CacheT<Arg>, DimsStatic<2, 1, 1>>;
+
     template <typename VUV, typename Pack, typename Arg>
     inline __device__ void operator()(VUV &vuv, bool isDiagonal, int coarse_x_cb, int coarse_parity, int i0, int j0, int parity, const Pack &pack, const Arg &arg)
     {
       using real = typename Arg::Float;
       using TileType = typename Arg::vuvTileType;
       const int dim_index = arg.dim_index % arg.Y_atomic.geometry;
-      __shared__ complex<storeType> X[Arg::max_color_height_per_block][Arg::max_color_width_per_block][4][Arg::coarseSpin][Arg::coarseSpin];
-      __shared__ complex<storeType> Y[Arg::max_color_height_per_block][Arg::max_color_width_per_block][4][Arg::coarseSpin][Arg::coarseSpin];
+      Cache<Arg> cache;
+      auto &X = cache.data()[0];
+      auto &Y = cache.data()[1];
 
       int x_ = coarse_x_cb % arg.aggregates_per_block;
       int tx = virtualThreadIdx(arg);
@@ -1416,7 +1423,7 @@ namespace quda {
         }
       }
 
-      __syncthreads();
+      cache.sync();
 
 #pragma unroll
       for (int i = 0; i < TileType::M; i++) {
@@ -1445,7 +1452,7 @@ namespace quda {
         }
       }
 
-      __syncthreads();
+      cache.sync();
 
       if (tx < Arg::coarseSpin*Arg::coarseSpin && (parity == 0 || arg.parity_flip == 1) ) {
 

include/kernels/color_spinor_pack.cuh

@@ -172,17 +172,24 @@ namespace quda {
   };
 
   template <> struct site_max<true> {
+    template <typename Arg> struct CacheDims {
+      static constexpr int Ms = spins_per_thread<true>(Arg::nSpin);
+      static constexpr int Mc = colors_per_thread<true>(Arg::nColor);
+      static constexpr int color_spin_threads = (Arg::nSpin / Ms) * (Arg::nColor / Mc);
+      static constexpr dim3 dims(dim3 block)
+      {
+        // pad the shared block size to avoid bank conflicts for native ordering
+        if (Arg::is_native) block.x = ((block.x + device::warp_size() - 1) / device::warp_size()) * device::warp_size();
+        block.y = color_spin_threads; // state the y block since we know it at compile time
+        return block;
+      }
+    };
+
     template <typename Arg> __device__ inline auto operator()(typename Arg::real thread_max, Arg &)
     {
       using real = typename Arg::real;
-      constexpr int Ms = spins_per_thread<true>(Arg::nSpin);
-      constexpr int Mc = colors_per_thread<true>(Arg::nColor);
-      constexpr int color_spin_threads = (Arg::nSpin/Ms) * (Arg::nColor/Mc);
-      auto block = target::block_dim();
-      // pad the shared block size to avoid bank conflicts for native ordering
-      if (Arg::is_native) block.x = ((block.x + device::warp_size() - 1) / device::warp_size()) * device::warp_size();
-      block.y = color_spin_threads; // state the y block since we know it at compile time
-      SharedMemoryCache<real> cache(block);
+      constexpr int color_spin_threads = CacheDims<Arg>::color_spin_threads;
+      SharedMemoryCache<real, CacheDims<Arg>> cache;
       cache.save(thread_max);
       cache.sync();
       real this_site_max = static_cast<real>(0);

include/kernels/dslash_clover_helper.cuh

@@ -203,7 +203,7 @@ namespace quda {
 
       Mat A = arg.clover(x_cb, clover_parity, chirality);
 
-      SharedMemoryCache<half_fermion> cache(target::block_dim());
+      SharedMemoryCache<half_fermion> cache;
 
       half_fermion in_chi[n_flavor]; // flavor array of chirally projected fermion
 #pragma unroll

include/kernels/dslash_coarse.cuh

@@ -301,7 +301,7 @@ namespace quda {
   template <> struct dim_collapse<true> {
     template <typename T, typename Arg> __device__ __host__ inline void operator()(T &out, int dir, int dim, const Arg &arg)
     {
-      SharedMemoryCache<T> cache(target::block_dim());
+      SharedMemoryCache<T> cache;
       // only need to write to shared memory if not master thread
       if (dim > 0 || dir) cache.save(out);
 

include/kernels/dslash_domain_wall_m5.cuh

@@ -220,7 +220,7 @@ namespace quda
     if (mobius_m5::use_half_vector()) {
       // if using shared-memory caching then load spinor field for my site into cache
       typedef ColorSpinor<real, Arg::nColor, 4 / 2> HalfVector;
-      SharedMemoryCache<HalfVector> cache(target::block_dim());
+      SharedMemoryCache<HalfVector> cache;
 
       { // forwards direction
         constexpr int proj_dir = dagger ? +1 : -1;
@@ -271,7 +271,7 @@ namespace quda
     } else { // use_half_vector
 
       // if using shared-memory caching then load spinor field for my site into cache
-      SharedMemoryCache<Vector> cache(target::block_dim());
+      SharedMemoryCache<Vector> cache;
       if (shared) {
         if (sync) { cache.sync(); }
         cache.save(in);
@@ -377,7 +377,7 @@ namespace quda
     const auto inv = arg.inv;
 
     // if using shared-memory caching then load spinor field for my site into cache
-    SharedMemoryCache<Vector> cache(target::block_dim());
+    SharedMemoryCache<Vector> cache;
     if (shared) {
       // cache.save(arg.in(s_ * arg.volume_4d_cb + x_cb, parity));
       if (sync) { cache.sync(); }
@@ -436,7 +436,7 @@ namespace quda
     Vector out;
 
     if (mobius_m5::use_half_vector()) {
-      SharedMemoryCache<HalfVector> cache(target::block_dim());
+      SharedMemoryCache<HalfVector> cache;
 
       { // first do R
         constexpr int proj_dir = dagger ? -1 : +1;
@@ -495,7 +495,7 @@ namespace quda
         out += l.reconstruct(4, proj_dir);
       }
     } else { // use_half_vector
-      SharedMemoryCache<Vector> cache(target::block_dim());
+      SharedMemoryCache<Vector> cache;
       if (shared) {
         if (sync) { cache.sync(); }
         cache.save(in);

include/kernels/dslash_mobius_eofa.cuh

@@ -107,7 +107,7 @@ namespace quda
         using real = typename Arg::real;
         typedef ColorSpinor<real, Arg::nColor, 4> Vector;
 
-        SharedMemoryCache<Vector> cache(target::block_dim());
+        SharedMemoryCache<Vector> cache;
 
         Vector out;
         cache.save(arg.in(s * arg.volume_4d_cb + x_cb, parity));
@@ -185,7 +185,7 @@ namespace quda
         typedef ColorSpinor<real, Arg::nColor, 4> Vector;
 
         const auto sherman_morrison = arg.sherman_morrison;
-        SharedMemoryCache<Vector> cache(target::block_dim());
+        SharedMemoryCache<Vector> cache;
         cache.save(arg.in(s * arg.volume_4d_cb + x_cb, parity));
         cache.sync();
 

include/kernels/dslash_ndeg_twisted_clover.cuh

@@ -72,7 +72,7 @@ namespace quda
         // apply the chiral and flavor twists
         // use consistent load order across s to ensure better cache locality
         Vector x = arg.x(my_flavor_idx, my_spinor_parity);
-        SharedMemoryCache<Vector> cache(target::block_dim());
+        SharedMemoryCache<Vector> cache;
         cache.save(x);
 
         x.toRel(); // switch to chiral basis

include/kernels/dslash_ndeg_twisted_clover_preconditioned.cuh

@@ -91,7 +91,7 @@ namespace quda
 
         int chirality = flavor; // relabel flavor as chirality
 
-        SharedMemoryCache<HalfVector> cache(target::block_dim());
+        SharedMemoryCache<HalfVector> cache;
 
         enum swizzle_direction {
           FORWARDS = 0,

include/kernels/dslash_ndeg_twisted_mass_preconditioned.cuh

@@ -95,7 +95,7 @@ namespace quda
       }
 
       if (!dagger || Arg::asymmetric) { // apply A^{-1} to D*in
-        SharedMemoryCache<Vector> cache(target::block_dim());
+        SharedMemoryCache<Vector> cache;
         if (isComplete<mykernel_type>(arg, coord) && active) {
           // to apply the preconditioner we need to put "out" in shared memory so the other flavor can access it
           cache.save(out);

include/kernels/gauge_stout.cuh

@@ -6,6 +6,7 @@
 #include <su3_project.cuh>
 #include <kernels/gauge_utils.cuh>
 #include <kernel.h>
+#include <thread_local_cache.h>
 
 namespace quda
 {
@@ -134,8 +135,8 @@ namespace quda
       }
 
       Link U, Q;
-      SharedMemoryCache<Link> Stap(target::block_dim());
-      SharedMemoryCache<Link> Rect(target::block_dim(), sizeof(Link));
+      ThreadLocalCache<Link, 0, computeStapleRectangleOps> Stap;
+      ThreadLocalCache<Link, 0, decltype(Stap)> Rect; // offset by Stap type to ensure non-overlapping allocations
 
       // This function gets stap = S_{mu,nu} i.e., the staple of length 3,
       // and the 1x2 and 2x1 rectangles of length 5. From the following paper:

include/kernels/gauge_utils.cuh

@@ -19,6 +19,7 @@ namespace quda
   // matrix+matrix = 18 floating-point ops
   // => Total number of floating point ops per function call
   // dims * (2*18 + 4*198) = dims*828
+  using computeStapleOps = thread_array<int, 4>;
   template <typename Arg, typename Staple, typename Int>
   __host__ __device__ inline void computeStaple(const Arg &arg, const int *x, const Int *X, const int parity, const int nu, Staple &staple, const int dir_ignore)
   {
@@ -94,6 +95,7 @@ namespace quda
   // matrix+matrix = 18 floating-point ops
   // => Total number of floating point ops per function call
   // dims * (8*18 + 28*198) = dims*5688
+  using computeStapleRectangleOps = thread_array<int, 4>;
   template <typename Arg, typename Staple, typename Rectangle, typename Int>
   __host__ __device__ inline void computeStapleRectangle(const Arg &arg, const int *x, const Int *X, const int parity, const int nu,
                                                          Staple &staple, Rectangle &rectangle, const int dir_ignore)

include/kernels/gauge_wilson_flow.cuh

@@ -4,6 +4,7 @@
 #include <kernels/gauge_utils.cuh>
 #include <su3_project.cuh>
 #include <kernel.h>
+#include <thread_local_cache.h>
 
 namespace quda
 {
@@ -71,8 +72,8 @@ namespace quda
       // This function gets stap = S_{mu,nu} i.e., the staple of length 3,
       // and the 1x2 and 2x1 rectangles of length 5. From the following paper:
       // https://arxiv.org/abs/0801.1165
-      SharedMemoryCache<Link> Stap(target::block_dim());
-      SharedMemoryCache<Link> Rect(target::block_dim(), sizeof(Link)); // offset to ensure non-overlapping allocations
+      ThreadLocalCache<Link, 0, computeStapleRectangleOps> Stap;
+      ThreadLocalCache<Link, 0, decltype(Stap)> Rect; // offset by Stap type to ensure non-overlapping allocations
       computeStapleRectangle(arg, x, arg.E, parity, dir, Stap, Rect, Arg::wflow_dim);
       Z = arg.coeff1x1 * static_cast<const Link &>(Stap) + arg.coeff2x1 * static_cast<const Link &>(Rect);
       break;