warp-lang 1.7.0__py3-none-manylinux_2_34_aarch64.whl

warp/native/tile.h

@@ -0,0 +1,2584 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include "builtin.h"
+
+#ifdef __clang__
+// disable warnings related to C++17 extensions on CPU JIT builds
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wc++17-extensions"
+#endif // __clang__
+
+// Check if the CUDA toolkit is available
+#if WP_ENABLE_CUDA || defined(__CUDACC_RTC__)
+
+    // If NVRTC is being used, do not include extra headers (NVRTC has built-in float4)
+    #ifdef __CUDACC_RTC__
+        // NVRTC: Use built-in float4 (no need for extra definitions)
+    #else
+        // NVCC: Include vector_types.h to get float4
+        #include <cuda_runtime.h>
+    #endif
+
+#else
+    // If CUDA is not available (e.g., macOS build), manually define float4
+    struct alignas(16) float4 {
+        float x, y, z, w;
+    };
+#endif 
+
+// only used while building the warp core library
+#ifndef WP_TILE_BLOCK_DIM
+#define WP_TILE_BLOCK_DIM 256
+#endif
+
+#if !defined(__CUDA_ARCH__)
+#define WP_TILE_SHARED static
+#define WP_TILE_SYNC void
+
+#else
+#define WP_TILE_SHARED __shared__
+#define WP_TILE_SYNC __syncthreads
+#endif
+
+#if defined(__CUDA_ARCH__) && !defined(__INTELLISENSE__)
+  #if defined(__CUDACC_RTC__) || (defined(__clang__) && defined(__CUDA__))
+    #define WP_PRAGMA_UNROLL _Pragma("unroll")
+    #define WP_PRAGMA_NO_UNROLL _Pragma("unroll 1")
+  #else
+    #define WP_PRAGMA_UNROLL #pragma unroll
+    #define WP_PRAGMA_NO_UNROLL #pragma unroll 1
+  #endif
+
+#else
+
+    #define WP_PRAGMA_UNROLL
+    #define WP_PRAGMA_NO_UNROLL
+
+#endif
+
+#define WP_USE_ASYNC_PIPELINE 0
+#define WP_USE_REGISTER_GEMM 0
+
+#if defined(__CUDACC_RTC__)
+#define WP_TILE_THREAD_IDX threadIdx.x
+#else
+#define WP_TILE_THREAD_IDX 0 
+#endif // 
+
+ 
+
+/* Tile Expressions
+
+[ ] Tiles
+    [x] Register, Shared, Global
+    [ ] Layouts
+        [x] Simple
+        [ ] Cute
+    [x] Remove Alloc type from tile_shared_t
+    [x] wp.launch_tiled() helper
+[ ] Creation
+    [x] zeros
+    [x] ones
+    [x] arange
+    [x] tile()
+    [x] untile()
+    [ ] fromfunction()
+    [ ] explicit storage
+[ ] Load/Store
+    [ ] 1D load/store variants
+    [ ] max_coord option for non-aligned loads
+    [ ] Indexed load
+    [x] wp.tile_atomic_add()
+[ ] Maps
+    [x] Support user functions
+    [x] Support built-in functions
+    [ ] Support for lambda functions
+    [ ] Infer tile_map() output from operator type (e.g.: dot for each element)
+[ ] Reductions
+    [x] Sum
+        [x] Forward
+        [x] Reverse
+    [x] Min
+    [x] Max
+    [x] Custom
+[x] MatMul
+    [x] Forward
+    [x] Reverse
+[ ] Operators
+    [ ] +, -, *, /, @?
+    [ ] += for matmul, e.g.: c += a@b, or c = a@b
+[ ] Reshape
+    [ ] Broadcasting
+    [ ] Transpose
+        [x] Shared
+        [ ] Register
+    [ ] Slice
+[ ] Runtime
+    [x] Compile-time block dimensions
+    [x] Switch between SIMT / Tile based execution if `block_dim` not provided to wp.launch()
+[ ] Examples
+    [ ] Point registration
+    [ ] GEMM
+    [ ] MLP
+    [ ] LayerNorm  
+    [ ] SoftMax
+    [ ] GEMM
+    [ ] warp.sim (CRBA)
+    [ ] Batched MLP
+    [ ] Layer norm
+    [ ] FNO + Burgers equation
+    [ ] Stochastic financial modeling
+    [ ] Convolution: https://github.com/NVIDIA/MinkowskiEngine/blob/master/src/convolution_kernel.cu#L123
+    [ ] MeshCNN (Modulus, Oliver)
+    [ ] BioNemo (Ali)
+    [ ] Skinning (David/Or/Vismay)
+    [ ] warp.sim (VBD)
+[ ] Error checking
+    [ ] Ensure functions passed to tile_map() are compatible with tile type
+    [ ] Ensure that args passed to tile ops are compatible
+    [ ] Ensure tile load/store operations don't go out of bounds of arrays in debug mode
+
+*/
+
+/*
+Notes on shared memory synchronization
+======================================
+
+Currently operations that write to shared memory tiles (e.g.: tile_load())
+must synchronize before they return through WP_TILE_SYNC(), this
+ensures subsequent read operations from the tile do not cause a race condition.
+
+For tile_shared_t adjoints, the gradient accumulation is done through shared
+memory atomics, i.e.: atomic_add(), since for broadcast tiles multiple threads
+may map to the same location. Synchronization is still required after these 
+updates, since subsequent operations e.g.: adj_tile_load() will store the
+gradients to memory, and all updates must be visible at that point, e.g.:
+
+    a = wp.tile_load(...)
+    b = wp.tile_load(...)
+    c = wp.tile_matmul(a, b)
+    wp.tile_store(c)
+
+    // loads incoming adjoints from global -> shared
+    wp.adj_tile_store(c, adj_c)
+    // consumes adj_c, requires synchronization
+    wp.adj_tile_matmul(a, b, adj_a, adj_b, adj_c)
+    // consumes adj_b, requires synchronization
+    wp.adj_tile_load(..., adj_b)
+    // consumes adj_b, requires synchronization
+    wp.adj_tile_load(..., adj_a)
+
+Generally synchronization to adjoint tiles will happen through the
+tile_shared_t::add() and tile_shared_t::assign() function automatically,
+but in some cases e.g.: tile_matmul() it is done manually.
+
+The current synchronization strategy is conservative, and can lead to more
+synchronization than necessary. A more sophisticated strategy would be
+to track the 'dirty' state of shared tiles, and synchronize only when
+necessary. In addition, custom synchronization for e.g.: tile_load()
+operations could be added through a SyncProvider template parameter on
+the tile_shared_t type, for example to support barrier synchronization
+for asynchronous global to shared loads.
+*/
+
+namespace wp
+{
+
+// Primary template
+template <typename T, typename U>
+struct is_same {
+    static constexpr bool value = false;
+};
+
+// Specialization for the case when T and U are the same type
+template <typename T>
+struct is_same<T, T> {
+    static constexpr bool value = true;
+};
+
+
+template <int N>
+struct tile_coord_t
+{
+    int indices[N];
+
+    CUDA_CALLABLE inline int operator[](int i) const { assert(0 <= 1 && i < N); return indices[i]; }
+    CUDA_CALLABLE inline int& operator[](int i) { assert(0 <= 1 && i < N); return indices[i]; }
+
+    CUDA_CALLABLE inline tile_coord_t<N> operator + (const tile_coord_t<N>& c) const
+    {
+        tile_coord_t<N> out;
+        for (int i=0; i < N; ++i)
+        {
+            out.indices[i] = indices[i] + c.indices[i];
+        }
+        return out;
+    }    
+};
+
+// This function deduces N = sizeof...(Ints)
+template <typename... Ints>
+constexpr tile_coord_t<sizeof...(Ints)> tile_coord(Ints... idxs)
+{
+    constexpr int N = sizeof...(Ints);
+    
+    // Create the result
+    tile_coord_t<N> result{};
+    
+    // Capture all arguments in a local array
+    int arr[] = { static_cast<int>(idxs)... };
+
+    // C++14 or later: 'for' is allowed in a constexpr context
+    for (int i = 0; i < N; ++i)
+    {
+        result.indices[i] = arr[i];
+    }
+
+    return result;
+}
+
+// helpers to construct a coord from a set of indices
+inline auto tile_coord(int i) 
+{
+    auto c = tile_coord_t<1>();
+    c.indices[0] = i;
+    return c;
+}
+
+inline auto tile_coord(int i, int j)
+{
+    auto c = tile_coord_t<2>();
+    c.indices[0] = i;
+    c.indices[1] = j;
+    return c;
+}
+
+inline auto tile_coord(int i, int j, int k)
+{
+    auto c = tile_coord_t<3>();
+    c.indices[0] = i;
+    c.indices[1] = j;
+    c.indices[2] = k;
+    return c;
+}    
+
+inline auto tile_coord(int i, int j, int k, int l)
+{
+    auto c = tile_coord_t<4>();
+    c.indices[0] = i;
+    c.indices[1] = j;
+    c.indices[2] = k;
+    c.indices[3] = l;
+    return c;
+}
+
+// represents a compile time int tuple for strides/shapes/coords
+template <int... V>
+struct tile_tuple_t
+{
+    static constexpr int N = sizeof...(V);
+    static_assert(N > 0, "Expected N > 0");
+
+    static constexpr int data[N] = { V... };
+
+    static constexpr int dim(int i) { assert(i < N); return data[i]; }
+    static constexpr int size() 
+    {
+        int res = data[0];
+        for (int i=1; i < N; ++i)
+            res *= data[i];
+
+        return res;
+    }
+};
+
+// simple helper to compute strides from a shape up to 4d
+template <typename Shape>
+struct compute_strides;
+
+// 1D
+template <int D0>
+struct compute_strides< tile_tuple_t<D0> > { using Stride = tile_tuple_t<1>; };
+// 2D
+template <int D0, int D1>
+struct compute_strides< tile_tuple_t<D0, D1> > { using Stride = tile_tuple_t<D1, 1>; };
+// 3D
+template <int D0, int D1, int D2>
+struct compute_strides< tile_tuple_t<D0, D1, D2> > { using Stride = tile_tuple_t<(D1 * D2), D2, 1>; };
+// 4D
+template <int D0, int D1, int D2, int D3>
+struct compute_strides< tile_tuple_t<D0, D1, D2, D3> > { using Stride = tile_tuple_t<(D1 * D2 * D3), (D2 * D3), D3, 1>; };
+
+
+// alias of tuple to represent shapes
+template <int... V>
+using tile_shape_t = tile_tuple_t<V...>;
+
+// alias of tuple to represent stride
+template <int... V>
+using tile_stride_t = tile_tuple_t<V...>;
+
+
+// represents a tile stored in global memory with dynamic strides
+// used to represent the source and offset for tile loads to register/shared
+template <typename T, typename Shape_>
+struct tile_global_t 
+{
+    using Type = T;
+    using Shape = Shape_;
+    using Coord = tile_coord_t<Shape::N>;
+
+    array_t<T> data;
+    Coord offset;
+    
+    tile_global_t(array_t<T>& a, const Coord& c) : data(a), offset(c)
+    {
+    }
+
+    inline CUDA_CALLABLE int index_from_coord(const Coord& coord) const
+    {
+        // element index
+        int index = 0;
+        
+        WP_PRAGMA_UNROLL
+        for (int i=0; i < Shape::N; ++i)
+        {
+            // global = offset + coord
+            int c = offset[i] + coord[i];
+            index += data.strides[i]*c;
+        }   
+
+        return index/sizeof(T);
+    }    
+    
+    inline CUDA_CALLABLE bool index(const Coord& coord, int& out) const
+    {
+        // element index
+        int index = 0;
+        
+        WP_PRAGMA_UNROLL
+        for (int i=0; i < Shape::N; ++i)
+        {
+            // global = offset + coord
+            int c = offset[i] + coord[i];
+
+            // handle out of bounds case
+            if (c >= data.shape[i])
+                return false;
+            else
+                index += data.strides[i]*c;
+        }   
+
+        // array strides are in bytes so we convert to elements
+        out = index / sizeof(T);
+        return true;
+    }
+
+    inline CUDA_CALLABLE T load(const Coord& coord) const
+    {   
+        int i;
+        if (index(coord, i))
+            return data.data[i];
+        else
+            return T(0);
+    }
+
+    inline CUDA_CALLABLE T load_grad(const Coord& coord) const
+    {
+        int i;
+        if (index(coord, i))
+            return data.grad[i];
+        else
+            return T(0);
+    }    
+
+    inline CUDA_CALLABLE void store(const Coord& coord, const T& x) const
+    {   
+        int i;
+        if (index(coord, i))
+            data.data[i] = x;
+    }
+
+    inline CUDA_CALLABLE T atomic_add(const Coord& coord, const T& value) const
+    {
+        int i;
+        if (index(coord, i))
+            return wp::atomic_add(&data.data[i], value);
+        else
+            return T(0);
+    }
+
+    inline CUDA_CALLABLE T atomic_add_grad(const Coord& coord, const T& grad) const
+    {   
+        int i;
+        if (index(coord, i))
+            return wp::atomic_add(&data.grad[i], grad);
+        else
+            return T(0);
+    }
+};
+
+template <typename Shape_>
+struct tile_layout_register_t
+{
+    using Shape = Shape_;
+    using Coord = tile_coord_t<Shape::N>;
+
+    static constexpr int Size = Shape::size();
+    static constexpr int NumRegs = (Size + WP_TILE_BLOCK_DIM - 1) / WP_TILE_BLOCK_DIM;
+    static constexpr bool Aligned = Size%WP_TILE_BLOCK_DIM == 0;
+
+    static inline CUDA_CALLABLE int linear_from_register(int reg)
+    {
+        return WP_TILE_THREAD_IDX + reg*WP_TILE_BLOCK_DIM;
+    }
+
+    static inline CUDA_CALLABLE int linear_from_coord(Coord c)
+    {
+        int linear = 0;
+        int stride = 1;
+        
+        WP_PRAGMA_UNROLL
+        for (int i=Shape::N-1; i >= 0; --i)
+        {
+            linear += c[i] * stride;
+            stride *= Shape::dim(i);
+        }
+        return linear;
+    }
+
+    static inline CUDA_CALLABLE auto coord_from_linear(int linear)
+    {
+        Coord c;
+        
+        WP_PRAGMA_UNROLL
+        for (int i=Shape::N-1; i >= 0; --i)
+        {
+            c[i] = linear%Shape::dim(i);
+            linear /= Shape::dim(i);
+        }
+
+        return c;
+    }
+    
+    static inline CUDA_CALLABLE int thread_from_linear(int linear)
+    {
+        const int thread = linear%WP_TILE_BLOCK_DIM;
+        return thread;
+    }
+
+    static inline CUDA_CALLABLE int register_from_linear(int linear)
+    {
+        const int reg = linear/WP_TILE_BLOCK_DIM;
+        return reg;
+    }
+
+    static inline CUDA_CALLABLE bool valid(int linear)
+    {
+        if (Aligned || linear < Size)
+            return true;
+        else
+            return false;
+    }
+
+};
+
+// represents a tile stored in registers across a block
+template <typename T, typename L>
+struct tile_register_t
+{
+    using Type = T;
+    using Layout = L;
+
+    T data[Layout::NumRegs];
+   
+    inline CUDA_CALLABLE tile_register_t(T value=T(0.0)) 
+    {
+        // zero-initialize by default necessary for tile adjoints
+        // need to check if this results in worse codegen
+        // than doing adj_var = tile_zeros() explicitly
+        // in backwards pass and letting default constructor
+        // avoid initialization
+        
+        for (int i=0; i < Layout::NumRegs; ++i)
+            data[i] = value;
+    }
+
+    inline CUDA_CALLABLE auto& operator=(const tile_global_t<T, typename Layout::Shape>& t)
+    {
+        copy_from_global(t);
+        return *this;
+    }
+
+    // define the += operator which is used during backward pass codegen
+    // when returning a register tile from a user defined function
+    inline CUDA_CALLABLE auto& operator += (tile_register_t<T, Layout>& rhs) 
+    {
+        grad_add(rhs);
+        return *this;
+    }
+
+    inline CUDA_CALLABLE T& operator()(int reg)
+    {
+        assert(reg < Layout::NumRegs);
+        return data[reg];
+    }    
+
+    inline CUDA_CALLABLE const T& operator()(int reg) const
+    {
+        assert(reg < Layout::NumRegs);
+        return data[reg];
+    }
+
+    inline CUDA_CALLABLE void assign(const tile_register_t<T, Layout>& tile) 
+    { 
+        for (int i=0; i < Layout::NumRegs; ++i)
+            data[i] = tile.data[i];
+    }
+
+    inline CUDA_CALLABLE void zero()
+    {
+        for (int i=0; i < Layout::NumRegs; ++i)
+            data[i] = T(0);
+    }
+
+    // extract a single tile element to a native type
+    template <typename Coord>
+    inline CUDA_CALLABLE Type extract(const Coord& c)
+    {
+        // map from logical coords (i, j) -> (thread, reg)
+        const int linear = Layout::linear_from_coord(c);
+        const int thread = Layout::thread_from_linear(linear);
+        const int reg = Layout::register_from_linear(linear);
+
+        WP_TILE_SHARED Type scratch;
+
+        // ensure any previously scheduled threads have finished reading from scratch
+        WP_TILE_SYNC();
+
+        if (WP_TILE_THREAD_IDX == thread)
+        {
+            scratch = data[reg];
+        }
+
+        // ensure extraction thread has updated smem
+        WP_TILE_SYNC();
+
+        return scratch;
+    }
+    
+
+    // backward version of scalar extract
+    template <typename Coord>
+    inline CUDA_CALLABLE void adj_extract(const Coord& c, Type adj_ret)
+    {
+        // map from logical coords (i, j) -> (thread, reg)
+        const int linear = Layout::linear_from_coord(c);
+        const int thread = Layout::thread_from_linear(linear);
+        const int reg = Layout::register_from_linear(linear);
+
+        if (WP_TILE_THREAD_IDX == thread)
+        {
+            data[reg] += adj_ret;
+        }
+    }
+
+    inline CUDA_CALLABLE void print() const;
+
+
+    // return the in-register version of this tile (nop)
+    inline CUDA_CALLABLE auto& copy_to_register() 
+    {
+        return *this; 
+    }
+
+    inline CUDA_CALLABLE const auto& copy_to_register() const 
+    {
+        return *this; 
+    }
+
+    // apply a lambda to all valid entries in the tile
+    // Op should be a functor that takes a register index and tile_coord_t as input
+    template <typename Op>
+    void apply(Op op)
+    {
+        WP_PRAGMA_UNROLL
+        for (int i=0; i < Layout::NumRegs; ++i)
+        {  
+            int linear = Layout::linear_from_register(i);
+            if (!Layout::valid(linear))
+                break;
+
+            auto c = Layout::coord_from_linear(linear);
+            op(i, c);
+        }        
+    }
+    
+
+    // in-place gradient zero
+    inline CUDA_CALLABLE void grad_zero()
+    {
+        zero();
+    }
+
+    // accumulate gradients onto this tile
+    inline CUDA_CALLABLE void grad_add(const tile_register_t<T, Layout>& tile) 
+    { 
+        for (int i=0; i < Layout::NumRegs; ++i)
+            data[i] += tile.data[i];
+    }
+
+    CUDA_CALLABLE void grad_add(const tile_global_t<T, typename Layout::Shape>& global) 
+    {
+        apply([&](int reg, auto c) {data[reg] = global.load_grad(c);});
+
+    }
+
+    inline CUDA_CALLABLE auto& grad_to_register() 
+    {
+        // nop for register tiles
+        return *this;
+    }
+
+    template <typename Global>
+    inline CUDA_CALLABLE void copy_to_global(const Global& dest)
+    {
+        apply([&](int reg, auto c) { dest.store(c, data[reg]); });
+    }
+
+    template <typename Global>
+    inline CUDA_CALLABLE void copy_from_global(const Global& src)
+    {
+        apply([&](int reg, auto c) { data[reg] = src.load(c); });
+    }
+
+    // add a register tile to a global array
+    template <typename Global>
+    inline CUDA_CALLABLE auto atomic_add(const Global& dest)
+    {
+        // allocate a tile to hold previous dest value
+        auto previous = *this;
+
+        apply([&](int reg, auto c) { previous.data[reg] = dest.atomic_add(c, data[reg]); });
+        return previous;
+    }
+
+    // add a register tile to the gradient of a global array
+    template <typename Global>
+    inline CUDA_CALLABLE auto atomic_add_grad(const Global& dest)
+    {
+        // allocate a tile to hold previous dest value
+        auto previous = *this;
+
+        apply([&](int reg, auto c) { previous.data[reg] = dest.atomic_add_grad(c, data[reg]); });
+        return previous;
+    }        
+};
+
+
+// helper to allocate a register tile like another tile
+// users can either specify a template explicitly or
+// pass in another concrete instance
+template<typename Tile>
+auto tile_register_like(Tile* t=nullptr)
+{
+    using T = typename Tile::Type;
+    using L = typename Tile::Layout;
+
+    return tile_register_t<T, tile_layout_register_t<typename L::Shape>>(T(0.0));
+}
+
+// helper to construct a register tile from a type and a list of dims
+template <typename T, int... Dims>
+auto tile_register()
+{
+    return tile_register_t<T, tile_layout_register_t<tile_shape_t<Dims...>>>();
+}
+
+inline CUDA_CALLABLE int tile_align(int num_bytes)
+{
+    // note this much match value in Python types.py
+    const int alignment = 16;
+
+    const int num_bytes_abs = num_bytes < 0 ? - num_bytes : num_bytes;
+    const int sign = num_bytes < 0 ? - 1 : 1;
+
+    return sign * ((num_bytes_abs + alignment - 1) / alignment) * alignment;
+}
+
+inline CUDA_CALLABLE void* tile_alloc_shared(int num_bytes, bool init=false, bool check=false)
+{
+    // we maintain a per-thread offset into dynamic
+    // shared memory that allows us to keep track of 
+    // current use across dynamic function calls
+    WP_TILE_SHARED int smem_base[WP_TILE_BLOCK_DIM];
+
+    if (init)
+    {
+        smem_base[WP_TILE_THREAD_IDX] = 0;
+        return nullptr;
+    }
+    else if (check)
+    {
+        assert(smem_base[WP_TILE_THREAD_IDX] == 0);
+        return nullptr;
+    }
+    else
+    {
+        const int offset = smem_base[WP_TILE_THREAD_IDX];
+        
+        // one entry per-thread so no need for synchronization
+        smem_base[WP_TILE_THREAD_IDX] += tile_align(num_bytes);
+
+#ifdef __CUDA_ARCH__
+        extern __shared__ char dynamic_smem_base[];
+#else
+        // on CPU allocate a fixed 256k block to use for shared allocs
+        static const int max_cpu_shared = 256*1024;
+        static char dynamic_smem_base[max_cpu_shared];
+
+        assert(smem_base[WP_TILE_THREAD_IDX] <= max_cpu_shared);
+#endif
+        return &(dynamic_smem_base[offset]);
+    }
+}
+
+
+template <typename Shape_, typename Stride_= typename compute_strides<Shape_>::Stride>
+struct tile_layout_strided_t
+{
+    using Shape = Shape_;
+    using Stride = Stride_;
+    using Coord = tile_coord_t<Shape::N>;
+    
+    static constexpr int Size = Shape::size();
+    static constexpr bool Aligned = Size%WP_TILE_BLOCK_DIM == 0;
+
+    static inline CUDA_CALLABLE auto coord_from_linear(int linear)
+    {
+        assert(linear < Size);
+
+        Coord c;
+        
+        WP_PRAGMA_UNROLL
+        for (int d=Shape::N-1; d >= 0; --d)
+        {
+            c[d] = linear%Shape::dim(d);
+            linear /= Shape::dim(d);
+        }
+
+        return c;
+    }
+
+    static inline CUDA_CALLABLE int index_from_coord(Coord c)
+    {
+        int index = 0;
+
+        WP_PRAGMA_UNROLL
+        for (int d=0; d < Shape::N; ++d)
+        {
+            assert(c[d] < Shape::dim(d));
+
+            index += c[d]*Stride::dim(d);
+        }
+
+        return index;
+    }
+
+    // checks whether a strided layout is unique, i.e.: if memory locations are only
+    // every referred to by one element in the tile, this is a basic test that only
+    // checks for broadcast dimensions, it would be possible to do the full check
+    // using sorted shape/strides in Python and add it as a template parameter to the type
+    static constexpr bool is_unique() 
+    {
+        constexpr int N = Shape::N;
+
+        // check for any broadcast dimensions
+        for (int i=0; i < N; ++i)
+            if (Stride::dim(i) == 0)
+                return false;
+        
+        return true;
+    }
+
+    static constexpr bool Unique = is_unique();
+
+    static inline CUDA_CALLABLE bool valid(int linear)
+    {
+        return linear < Size;
+    }    
+
+};
+
+
+template <typename T, typename L, bool Owner_=true>
+struct tile_shared_t 
+{
+    using Type = T;
+    using Layout = L;
+    static constexpr bool Owner = Owner_;
+
+    struct Storage
+    {
+        T* ptr;
+
+        Storage(T* p) : ptr(p) {}
+
+        inline CUDA_CALLABLE T& operator()(typename Layout::Coord c)
+        {
+            assert(ptr);
+
+            int index = Layout::index_from_coord(c);
+            return ptr[index];
+        }
+
+        inline CUDA_CALLABLE const T& operator()(typename Layout::Coord c) const
+        {          
+            assert(ptr);
+
+            int index = Layout::index_from_coord(c);
+            return ptr[index];
+        }
+
+        inline CUDA_CALLABLE T& operator()(int linear)
+        {
+            assert(ptr);
+            assert(Layout::valid(linear));
+
+            auto c = Layout::coord_from_linear(linear);
+            return (*this)(c);
+        }    
+
+        inline CUDA_CALLABLE const T& operator()(int linear) const
+        {
+            assert(ptr);
+            assert(Layout::valid(linear));
+
+            auto c = Layout::coord_from_linear(linear);
+            return (*this)(c);
+        }            
+    };
+
+    Storage data;
+    Storage grad;
+
+    // we need to track whether or not this tile's data has been initialized.
+    // once true, any re-initialization of data that follows needs a WP_TILE_SYNC()
+    // call to precede it, to allow threads that are still reading from this tile
+    // to complete their work. e.g, in a dynamic loop:
+    // for i in range(x):
+    //     tile = wp.tile_load(arr, i, TILE_SIZE, storage="shared")
+    //     # read from tile...
+    bool initialized;
+
+    // default initialization (non-initialized)
+    inline CUDA_CALLABLE tile_shared_t() : data(nullptr), grad(nullptr), initialized(false)
+    {
+    }
+
+    // initialize from an existing tile's memory
+    inline CUDA_CALLABLE tile_shared_t(T* data, T* grad=nullptr, bool initialized=true) : data(data), grad(grad), initialized(initialized)
+    {
+    }
+
+    inline CUDA_CALLABLE ~tile_shared_t()
+    {
+        if (Owner)
+        {
+            // update our per-thread shared memory allocator
+            if (data.ptr)
+                tile_alloc_shared(-Layout::Size*int(sizeof(T)));
+
+            if (grad.ptr)
+                tile_alloc_shared(-Layout::Size*int(sizeof(T)));
+        }
+    }
+
+    // assign from a register tile
+    template <typename Tile>
+    inline CUDA_CALLABLE auto& operator=(const Tile& t)
+    {
+        assign(t);
+        return *this;
+    }
+
+
+/*
+    // construct from another shared tile, this constructor
+    // is invoked for reshape operations like `wp.tile_transpose()`
+    template <typename OtherT, typename OtherLayout>
+    inline CUDA_CALLABLE auto& operator=(const tile_shared_t<OtherT, OtherLayout>& rhs) 
+    {
+        using OtherTile = tile_shared_t<OtherT, OtherLayout>;
+
+        // check dimensions are compatible
+        static_assert(Size == OtherTile::Size, "Expected Size == OtherTile::Size");
+
+        // alias tile directly
+        data = rhs.data;
+        grad = rhs.grad;
+        initialized = rhs.initialized;
+
+        return *this;
+    }    
+*/
+
+    // assign from a global tile (load)
+    inline CUDA_CALLABLE auto& operator=(const tile_global_t<T, typename Layout::Shape>& t)
+    {        
+        copy_from_global(t);
+        return *this;
+    }
+
+    // assign from a constant value
+    inline CUDA_CALLABLE auto& operator=(const T& x)
+    {
+        // sync if we are re-initializing data so that any threads that are still
+        // reading from this tile can complete their work, e.g.: if re-assigning
+        // to a tile during a dynamic loop
+        if (initialized)
+            WP_TILE_SYNC();
+
+        for (int i=WP_TILE_THREAD_IDX; i < Layout::Size; i+= WP_TILE_BLOCK_DIM)
+            data(i) = x;
+
+        initialized = true;
+        WP_TILE_SYNC();
+        return *this;
+    }
+
+    // in-place zero
+    inline CUDA_CALLABLE void zero()
+    {
+        for (int i=WP_TILE_THREAD_IDX; i < Layout::Size; i+= WP_TILE_BLOCK_DIM)
+            data(i) = T(0);
+
+        WP_TILE_SYNC();
+    }
+
+    // extract a single tile element to a native type
+    inline CUDA_CALLABLE Type extract(const typename Layout::Coord& c)
+    {
+        return data(c);
+    }
+        
+    // backward of scalar extraction
+    inline CUDA_CALLABLE void adj_extract(const typename Layout::Coord& c, Type adj_ret)
+    {
+        // since multiple threads may extract the same element
+        // we need to accumulate using atomic operations
+        wp::atomic_add(&grad(c), adj_ret);
+    
+        WP_TILE_SYNC();       
+    }
+
+
+    // copy register tile to shared
+    template <typename Tile>
+    inline CUDA_CALLABLE void assign(const Tile& tile)
+    { 
+        if (initialized)
+            WP_TILE_SYNC();
+
+        WP_PRAGMA_UNROLL
+        for (int i=0; i < Tile::Layout::NumRegs; ++i)
+        {
+            const int linear = Tile::Layout::linear_from_register(i);
+
+            // handle case where tile size is not
+            // aligned to block dimensions
+            if (!Tile::Layout::valid(linear))
+                break;         
+
+            data(linear) = tile.data[i];
+        }
+
+        initialized = true;
+        WP_TILE_SYNC();
+    }
+
+    // in-place gradient zero
+    inline CUDA_CALLABLE void grad_zero()
+    {
+        for (int i=WP_TILE_THREAD_IDX; i < Layout::Size; i+= WP_TILE_BLOCK_DIM)
+            grad(i) = T(0);
+
+        WP_TILE_SYNC();
+    }
+
+
+    // accumulate gradients onto this tile
+    template <typename Tile>
+    inline CUDA_CALLABLE void grad_add(const Tile& tile) 
+    { 
+        WP_PRAGMA_UNROLL
+        for (int i=0; i < Tile::Layout::NumRegs; ++i)
+        {
+            const int linear = Tile::Layout::linear_from_register(i);
+
+            // handle case where tile size is not
+            // aligned to block dimensions
+            if (!Tile::Layout::valid(linear))
+                break;
+
+            // if the destination layout is unique (no broadcast dimensions)
+            // then we can use regular non-atomic accmulation
+            if (Layout::Unique)
+                grad(linear) += tile.data[i];
+            else
+                // use shared memory atomics to accumulate gradients
+                // since for broadcast tiles (e.g.: a bias vector) multiple incoming threads 
+                // may map to a single location in shared memory
+                wp::atomic_add(&grad(linear), tile.data[i]);
+            
+        }
+
+        WP_TILE_SYNC();
+    }
+
+    // accumulate gradient onto this tile from a global array
+    CUDA_CALLABLE void grad_add(const tile_global_t<T, typename Layout::Shape>& global) 
+    {
+        WP_PRAGMA_UNROLL
+        for (int i=WP_TILE_THREAD_IDX; i < Layout::Size; i += WP_TILE_BLOCK_DIM)
+        {  
+            auto c = Layout::coord_from_linear(i);
+            T g = global.load_grad(c);
+
+            if (Layout::Unique)
+            {
+                // if the destination layout is unique (no broadcast dimensions)
+                // then we can use regular non-atomic accumulation
+                grad(c) += g;
+            }
+            else
+            {
+                // use shared memory atomics to accumulate gradients
+                // since for broadcast tiles (e.g.: a bias vector) multiple incoming threads 
+                // may map to a single location in shared memory
+                wp::atomic_add(&grad(c), g);
+            }
+        }
+
+        WP_TILE_SYNC();        
+    }    
+
+    // copy shared tile to register
+    inline CUDA_CALLABLE auto grad_to_register() 
+    { 
+        using Tile = tile_register_t<T, tile_layout_register_t<typename Layout::Shape>>;
+        Tile out;
+
+        WP_PRAGMA_UNROLL
+        for (int i=0; i < Tile::Layout::NumRegs; ++i)
+        {
+            const int linear = Tile::Layout::linear_from_register(i);
+
+            if (!Tile::Layout::valid(linear))
+                break;
+
+            out(i) = grad(linear);
+        }
+
+        return out;
+    }
+
+    // copy shared tile to register
+    inline CUDA_CALLABLE auto copy_to_register() const
+    { 
+
+        auto out = tile_register_like(this);
+
+        using Layout = typename decltype(out)::Layout;
+
+        WP_PRAGMA_UNROLL
+        for (int i=0; i < Layout::NumRegs; ++i)
+        {
+            const int linear = Layout::linear_from_register(i);
+
+            if (!Layout::valid(linear))
+                break;
+
+            out(i) = data(linear);
+        }
+
+        return out;
+    }
+
+    template <typename Global>
+    inline CUDA_CALLABLE void copy_to_global(const Global& dest)
+    {       
+
+#if defined(__CUDA_ARCH__)
+        // vectorized loads for specific input/output shapes
+        if constexpr (Layout::Shape::N == 2)
+        {
+            constexpr int lastdim = Layout::Shape::N-1;
+            constexpr bool contiguous_src = Layout::Stride::dim(lastdim) == 1;
+            const bool contiguous_dest = dest.data.strides[lastdim] == sizeof(T);
+            const int elements = (dest.data.shape[lastdim] - dest.offset[lastdim]);
+            const bool aligned_size = (elements*sizeof(T))%sizeof(float4) == 0;
+           
+            float4* dest128 = (float4*)&dest.data.data[dest.index_from_coord(tile_coord(0,0))];
+            const bool aligned_dst = (uint64_t)(dest128)%sizeof(float4) == 0;
+
+            if (contiguous_dest && contiguous_src && aligned_size && aligned_dst)
+            {                    
+                constexpr int M = Layout::Shape::dim(0);
+                constexpr int N = (Layout::Shape::dim(1)*sizeof(T))/sizeof(float4);            
+
+                // alias of shared tile with 128bit type
+                using SrcLayout = tile_layout_strided_t<tile_shape_t<M, N>>;
+                tile_shared_t<float4, SrcLayout> src128((float4*)data.ptr);
+
+                assert(((uint64_t)(data.ptr))%sizeof(float4) == 0);
+                assert(((uint64_t)(dest128))%sizeof(float4) == 0);
+
+                const int stride_i = dest.data.strides[0]/sizeof(float4);
+                const int stride_j = 1;
+
+                WP_PRAGMA_UNROLL
+                for (int i=WP_TILE_THREAD_IDX; i < SrcLayout::Size; i += WP_TILE_BLOCK_DIM)
+                {  
+                    auto c = SrcLayout::coord_from_linear(i);
+                    
+                    dest128[stride_i*c[0] + stride_j*c[1]] = src128.data(i);
+                }
+
+                return;
+            }
+        }
+
+#endif //defined(__CUDA_ARCH__)
+
+        // scalar bounds checked path
+        WP_PRAGMA_UNROLL
+        for (int i=WP_TILE_THREAD_IDX; i < Layout::Size; i += WP_TILE_BLOCK_DIM)
+        {
+            auto c = Layout::coord_from_linear(i);
+            dest.store(c, data(i));
+        }
+    }
+
+    inline CUDA_CALLABLE void cp_async_global_to_shared_128(float4* shared_dest, const float4* global_src)
+    {
+    #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+
+        unsigned long long saddr = 0ULL;
+        unsigned long long gaddr = 0ULL;
+
+        asm volatile("cvta.to.shared.u64 %0, %1;" : "=l"(saddr) : "l"(shared_dest));
+        asm volatile("cvta.to.global.u64 %0, %1;" : "=l"(gaddr) : "l"(global_src));
+
+        // Use cp.async on newer architectures
+        asm volatile(
+            "cp.async.ca.shared.global [%0], [%1], 16;\n"
+            :
+            : "l"(saddr), "l"(gaddr)
+        );
+    #else
+        // use regular load/store through register on older arches
+        *shared_dest = *global_src;
+    #endif
+    }    
+
+    inline CUDA_CALLABLE void cp_async_commit_and_wait_all_128()
+    {
+    #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+        asm volatile(
+            "cp.async.commit_group;\n"
+            "cp.async.wait_group 0;\n" ::);
+    #endif
+    }
+
+    template <typename Global>
+    inline CUDA_CALLABLE void copy_from_global(const Global& src)
+    {   
+        if (initialized)
+            WP_TILE_SYNC();
+        
+#if defined(__CUDA_ARCH__)
+
+        // vectorized loads for specific input/output shapes
+        if constexpr (Layout::Shape::N == 2)
+        {
+            constexpr int lastdim = Layout::Shape::N-1;
+            constexpr bool contiguous_dest = Layout::Stride::dim(lastdim) == 1;
+            const bool contiguous_src = src.data.strides[lastdim] == sizeof(T);
+            const int elements = (src.data.shape[lastdim] - src.offset[lastdim]);
+            const bool aligned_size = (elements*sizeof(T))%sizeof(float4) == 0;
+
+            float4* src128 = (float4*)&src.data.data[src.index_from_coord(tile_coord(0,0))];
+            const bool aligned_src = (uint64_t)(src128)%sizeof(float4) == 0;
+
+            if (contiguous_dest && contiguous_src && aligned_size && aligned_src)
+            {            
+                constexpr int M = Layout::Shape::dim(0);
+                constexpr int N = (Layout::Shape::dim(1)*sizeof(T))/sizeof(float4);
+
+                // alias of shared tile with 128bit type
+                using DestLayout = tile_layout_strided_t<tile_shape_t<M, N>>;
+                tile_shared_t<float4, DestLayout> dest128((float4*)data.ptr);                
+
+                assert(((uint64_t)(dest128.data.ptr))%sizeof(float4) == 0);
+                assert(((uint64_t)(src128))%sizeof(float4) == 0);
+
+                const int stride_i = src.data.strides[0]/sizeof(float4);
+                const int stride_j = 1;
+
+                WP_PRAGMA_UNROLL
+                for (int i=WP_TILE_THREAD_IDX; i < DestLayout::Size; i += WP_TILE_BLOCK_DIM)
+                {  
+                    auto c = DestLayout::coord_from_linear(i);
+                    
+#if WP_USE_ASYNC_PIPELINE
+                    cp_async_global_to_shared_128(&dest128.data(i), &src128[stride_i*c[0] + stride_j*c[1]]);
+#else
+                    dest128.data(i) = src128[stride_i*c[0] + stride_j*c[1]];
+#endif // WP_USE_ASYNC_PIPELINE
+                }
+
+#if WP_USE_ASYNC_PIPELINE
+                cp_async_commit_and_wait_all_128();
+#endif // WP_USE_ASYNC_PIPELINE
+
+                initialized = true;
+                WP_TILE_SYNC();
+                return;
+            }
+        }
+
+#endif //defined(__CUDA_ARCH__)
+
+        // scalar bounds checked path
+        WP_PRAGMA_UNROLL
+        for (int i=WP_TILE_THREAD_IDX; i < Layout::Size; i += WP_TILE_BLOCK_DIM)
+        {  
+            auto c = Layout::coord_from_linear(i);
+            data(i) = src.load(c);
+        }
+
+        initialized = true;
+        WP_TILE_SYNC();
+    }
+
+    template <typename Global>
+    inline CUDA_CALLABLE auto atomic_add(Global& dest)
+    {
+        copy_to_register().atomic_add(dest);
+    }
+
+    template <typename Global>
+    inline CUDA_CALLABLE auto atomic_add_grad(Global& dest)
+    {
+        grad_to_register().atomic_add_grad(dest);
+    }
+
+    // overload for integral types
+    inline CUDA_CALLABLE void print_value(int x) const
+    {
+        printf("%d", x);
+    }
+
+    // overload for floating point types
+    template <typename ValueType>
+    inline CUDA_CALLABLE void print_value(ValueType x) const
+    {
+        printf("%g", x);
+    }
+
+    template <int Level = 0>
+    inline CUDA_CALLABLE void print_values(const Storage& storage, int index=0) const
+    {
+        using Shape = typename Layout::Shape;
+
+        if constexpr (Level < Shape::N)
+        {
+            if constexpr (Level == Shape::N - 1)
+            {
+                // Special handling for 1D case
+                printf("[");
+                for (int i = 0; i < Shape::dim(Level); ++i)
+                {
+                    print_value(storage(index + i));
+
+                    if (i < Shape::dim(Level) - 1)
+                    {
+                        printf(" ");
+                    }                    
+                }
+                printf("]");                    
+            }
+            else if constexpr (Level == Shape::N - 2)
+            {                   
+                // Special handling for 2D case
+                printf("[");
+                for (int i = 0; i < Shape::dim(Level); ++i)                
+                {             
+                    printf("[");
+                    for (int j=0; j < Shape::dim(Level+1); ++j)
+                    {
+                        print_value(storage(index));                        
+
+                        if (j < Shape::dim(Level+1) - 1)
+                        {
+                            printf(" ");
+                        }
+
+                        ++index;
+                    }    
+
+                    printf("]");
+
+                    // next row
+                    if (i < Shape::dim(Level)-1)
+                    {
+                        printf("\n");
+
+                        // indent next row
+                        for (int i=0; i <= Shape::N-2; ++i)
+                            printf(" ");
+
+                    }
+                }
+                printf("]");
+            } 
+            else
+            {
+                printf("[");
+                for (int i = 0; i < Shape::dim(Level); ++i)
+                {
+                    print_values<Level + 1>(storage, index + i * Shape::dim(Level));
+                    if (i < Shape::dim(Level) - 1)
+                    {
+                        printf("\n\n");
+
+                        // indent next row
+                        for (int i=0; i <= Level; ++i)
+                            printf(" ");
+                    }
+                }
+                printf("]");
+            }
+        }
+    }
+
+    inline CUDA_CALLABLE void print(bool reverse=false) const
+    {
+        if (WP_TILE_THREAD_IDX != 0)
+            return;
+
+        if (reverse)
+            print_values(grad);
+        else
+            print_values(data);
+
+        printf(" = tile(shape=(");
+        for (int i=0; i < Layout::Shape::N; ++i)
+        {
+            printf("%d", Layout::Shape::dim(i));
+            if (i != Layout::Shape::N-1)
+                printf(",");
+        }
+
+        printf("), storage=shared)\n");
+    }    
+};
+
+
+template <typename T, typename L>
+void tile_register_t<T, L>::print() const
+{
+    // create a temporary shared tile so that
+    // we can print it deterministically
+    WP_TILE_SHARED T smem[L::Size];
+    tile_shared_t<T, tile_layout_strided_t<typename L::Shape>, false> scratch(smem, nullptr);
+
+    scratch.assign(*this);
+
+    WP_TILE_SYNC();
+
+    if (WP_TILE_THREAD_IDX == 0)
+    {
+        scratch.print_values(scratch.data, 0);
+
+        printf(" = tile(shape=(");
+        for (int i=0; i < L::Shape::N; ++i)
+        {
+            printf("%d", L::Shape::dim(i));
+            if (i != L::Shape::N-1)
+                printf(",");
+        }
+
+        printf("), storage=register)\n");
+    }
+
+    WP_TILE_SYNC();
+}
+
+// print entry points
+template <typename T, typename L>
+inline CUDA_CALLABLE void print(const tile_register_t<T, L>& t) { t.print(); }
+template <typename T, typename L, bool Owner>
+inline CUDA_CALLABLE void print(const tile_shared_t<T, L, Owner>& t) { t.print(); }
+
+template <typename T, typename L, bool O>
+inline CUDA_CALLABLE int len(const tile_shared_t<T, L, O>& t)
+{
+    return L::Shape::dim(0);
+}
+
+template <typename T, typename L, bool O, typename AdjTile>
+inline CUDA_CALLABLE void adj_len(const tile_shared_t<T,L,O>& t, const AdjTile& a, int& adj_ret)
+{
+}
+
+template <typename T, typename L>
+inline CUDA_CALLABLE int len(const tile_register_t<T, L>& t)
+{
+    return L::Shape::dim(0);
+}
+
+template <typename T, typename L, typename AdjTile>
+inline CUDA_CALLABLE void adj_len(const tile_register_t<T,L>& t, const AdjTile& a, int& adj_ret)
+{
+}
+
+
+template <typename T, typename L>
+inline CUDA_CALLABLE void adj_print(const tile_register_t<T, L>& t, const tile_register_t<T, L>& a) { a.print(); }
+template <typename T, typename L, bool Owner>
+inline CUDA_CALLABLE void adj_print(const tile_shared_t<T, L, Owner>& t, const tile_shared_t<T, L, Owner>& a) { a.print(true); }
+
+
+
+// helpers to allocate shared tiles
+template <typename T, typename Shape, bool RequiresGrad>
+inline CUDA_CALLABLE auto tile_alloc_empty()
+
+{   constexpr int size = Shape::size();
+    T* data = (T*)tile_alloc_shared(size*sizeof(T));
+    T* grad = nullptr;
+
+#if FP_CHECK
+
+    // initialize tile to quiet nan
+    uint32_t qnanbits = 0x7FC00000;
+    float qnan = *(float*)(&qnanbits);
+
+    for (int i=WP_TILE_THREAD_IDX; i < size; i+= WP_TILE_BLOCK_DIM)
+        data[i] = T(qnan);
+
+    WP_TILE_SYNC();
+
+#endif // FP_CHECK
+
+
+    if (RequiresGrad)
+    {
+        grad = (T*)tile_alloc_shared(size*sizeof(T));
+
+        for (int i=WP_TILE_THREAD_IDX; i < size; i+= WP_TILE_BLOCK_DIM)
+            grad[i] = T(0);
+
+        WP_TILE_SYNC();
+    }
+       
+    return tile_shared_t<T, tile_layout_strided_t<Shape>>(data, grad);
+}
+
+
+//-----------------------------------------------------------------------------------------------------
+// High level entry points for each op (correspond to one Warp builtin)
+
+// construct a tile from a local SIMT value (one per-thread)
+template <typename T>
+inline CUDA_CALLABLE auto tile(const T& x)
+{
+    tile_register_t<T, tile_layout_register_t<tile_shape_t<WP_TILE_BLOCK_DIM>>> result;
+    
+    using Layout = typename decltype(result)::Layout;
+    static_assert(Layout::NumRegs == 1, "Expected Layout::NumRegs == 1");
+
+    result.data[0] = x;
+    return result;
+}
+
+// overload for constructing a tile from a per-thread vector
+template <typename T, unsigned Length>
+inline CUDA_CALLABLE auto tile(const wp::vec_t<Length, T>& x)
+{
+    tile_register_t<T, tile_layout_register_t<tile_shape_t<Length, WP_TILE_BLOCK_DIM>>> result;
+    
+    using Layout = typename decltype(result)::Layout;
+    static_assert(Layout::NumRegs == Length, "Expected Layout::NumRegs == Length");
+
+    for (int i=0; i < Length; ++i)
+        result.data[i] = x[i]; 
+
+    return result;
+}
+
+// construct a tile from a local SIMT value (one per-thread)
+template <typename T, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile(const T& x, T& adj_x, AdjTile& adj_ret)
+{
+    static_assert(AdjTile::Layout::Shape::N == 1, "Expected AdjTile::Layout::Shape::N == 1");
+    static_assert(AdjTile::Layout::Shape::dim(0) == WP_TILE_BLOCK_DIM, "Expected AdjTile::Layout::Shape::dim(0) == WP_TILE_BLOCK_DIM");
+    
+    auto adj_reg = adj_ret.copy_to_register();
+
+    adj_x += adj_reg.data[0];
+}
+
+template <typename T, unsigned Length, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile(const wp::vec_t<Length, T>& x, wp::vec_t<Length, T>& adj_x, AdjTile& adj_ret)
+{
+    static_assert(AdjTile::Layout::Shape::N == 2, "Expected AdjTile::Layout::Shape::N == 2");
+    static_assert(AdjTile::Layout::Shape::dim(0) == Length, "Expected AdjTile::Layout::Shape::dim(0) == Length");
+    static_assert(AdjTile::Layout::Shape::dim(1) == WP_TILE_BLOCK_DIM, "Expected AdjTile::Layout::Shape::dim(1) == WP_TILE_BLOCK_DIM");
+
+    auto adj_reg = adj_ret.copy_to_register();
+
+    for (int i=0; i < Length; ++i)
+        adj_x[i] += adj_reg.data[i];
+}
+
+template <typename Tile>
+inline CUDA_CALLABLE auto untile(Tile& tile)
+{    
+    // code-gen should have set the tile to 
+    // have exactly the block dimension so 
+    // there is exactly one value per-thread
+    auto reg = tile.copy_to_register();
+
+    constexpr int N = Tile::Layout::Shape::N;
+
+    // scalar case
+    if constexpr(N == 1)
+    {
+        return reg.data[0];
+    }
+        
+    // vector case
+    if constexpr(N == 2)
+    {
+        constexpr int Length = Tile::Layout::Shape::dim(0);
+        wp::vec_t<Length, typename Tile::Type> v;
+        for (int i=0; i < Length; ++i)
+            v[i] = reg.data[i];
+
+        return v;
+    }
+}
+
+template <typename Tile, typename Value>
+inline CUDA_CALLABLE void adj_untile(Tile& tile, Tile& adj_tile, Value& adj_ret)
+{    
+    auto adj = adj_tile.copy_to_register();   
+    
+    constexpr int N = Tile::Layout::Shape::N;
+
+    // scalar case
+    if constexpr(N == 1)
+    {
+        adj.data[0] += adj_ret;
+    }
+
+    // vector case
+    if constexpr(N == 2)
+    {
+        constexpr int Length = Tile::Layout::Shape::dim(0);
+        for (int i=0; i < Length; ++i)
+            adj.data[i] += adj_ret[i];
+    }
+
+    adj_tile.assign(adj);
+}
+
+// zero initialized tile
+template <typename T, unsigned... Shape>
+inline CUDA_CALLABLE auto tile_zeros()
+{
+    // tile variable assignment operator will handle initialization (since lhs could be shared/register tile)
+    return T(0);
+}
+
+// one-initialized tile
+template <typename T, unsigned... Shape>
+inline CUDA_CALLABLE auto tile_ones()
+{
+    // tile variable assignment operator will handle initialization (since lhs could be shared/register tile)
+    return T(1);
+}
+
+// tile with evenly spaced values
+template <typename T, int Len>
+inline CUDA_CALLABLE auto tile_arange(T start, T stop, T step)
+{
+    auto out = tile_register<T, Len>();
+
+    using Layout = typename decltype(out)::Layout;
+    
+    WP_PRAGMA_UNROLL
+    for (int i=0; i < Layout::NumRegs; ++i)
+    {
+        const int linear = Layout::linear_from_register(i);
+
+        // handle case where tile size is not
+        // aligned to block dimensions
+        if (!Layout::valid(linear))
+            break;
+
+        out.data[i] = start + linear*step;
+    }
+    
+    return out;
+}
+
+template <typename T, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_arange(T start, T stop, T step,
+                                          T& adj_start, T& adj_stop, T& adj_step, AdjTile& adj_ret) {}
+
+// entry point for load operations, these just return a reference to a global memory array + coordinate
+template <unsigned... Shape, typename... Indices, typename T>
+inline CUDA_CALLABLE auto tile_load(array_t<T>& src, Indices... offset) 
+{
+    return tile_global_t<T, tile_shape_t<Shape...>>(src, tile_coord(offset...)); 
+}
+
+// // entry point for tile store operations
+// template <typename... Indices, typename T, typename Tile>
+// inline CUDA_CALLABLE void tile_store(array_t<T>& dest, Tile& src, Indices... x)
+// {
+//     src.copy_to_global(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x))); 
+// }
+
+// entry point for tile store operations
+template <typename T, typename Tile>
+inline CUDA_CALLABLE void tile_store(array_t<T>& dest, int x, Tile& src) { src.copy_to_global(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x))); }
+template <typename T, typename Tile>
+inline CUDA_CALLABLE void tile_store(array_t<T>& dest, int x, int y, Tile& src) { src.copy_to_global(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x, y))); }
+template <typename T, typename Tile>
+inline CUDA_CALLABLE void tile_store(array_t<T>& dest, int x, int y, int z, Tile& src) { src.copy_to_global(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x, y, z))); }
+template <typename T, typename Tile>
+inline CUDA_CALLABLE void tile_store(array_t<T>& dest, int x, int y, int z, int w, Tile& src) { src.copy_to_global(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x, y, z, w))); }
+
+
+
+template <typename T, typename Tile>
+inline CUDA_CALLABLE auto tile_atomic_add(array_t<T>& dest, int x, Tile& src) { return src.atomic_add(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x))); }
+template <typename T, typename Tile>
+inline CUDA_CALLABLE auto tile_atomic_add(array_t<T>& dest, int x, int y, Tile& src) { return src.atomic_add(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x, y)));}
+template <typename T, typename Tile>
+inline CUDA_CALLABLE auto tile_atomic_add(array_t<T>& dest, int x, int y, int z, Tile& src) { return src.atomic_add(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x, y, z)));}
+template <typename T, typename Tile>
+inline CUDA_CALLABLE auto tile_atomic_add(array_t<T>& dest, int x, int y, int z, int w, Tile& src) { return src.atomic_add(tile_global_t<T, typename Tile::Layout::Shape>(dest, tile_coord(x, y, z, w)));}
+
+
+//-------------------------------------
+// Adjoints
+
+template <typename T, typename AdjTile, typename Coord>
+inline CUDA_CALLABLE void adj_tile_load(array_t<T>& src, Coord c,
+                                        array_t<T>& adj_src, Coord adj_c,
+                                        AdjTile& adj_ret)
+{
+    tile_global_t<T, typename AdjTile::Layout::Shape> dest(src, c);
+    
+    // we allow users to override grad of src
+    if (adj_src.data)
+        dest.data.grad = adj_src.data;
+
+    adj_ret.atomic_add_grad(dest);   
+}
+
+
+template <typename T, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_load(array_t<T>& src, int x, array_t<T>& adj_src, int adj_x, AdjTile& adj_ret) { adj_tile_load( src, tile_coord(x), adj_src, tile_coord(0), adj_ret); }
+template <typename T, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_load(array_t<T>& src, int x, int y, array_t<T>& adj_src, int adj_x, int adj_y, AdjTile& adj_ret) { adj_tile_load( src, tile_coord(x, y), adj_src, tile_coord(0,0), adj_ret); }
+template <typename T, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_load(array_t<T>& src, int x, int y, int z, array_t<T>& adj_src, int adj_x, int adj_y, int adj_z, AdjTile& adj_ret) { adj_tile_load( src, tile_coord(x, y, z), adj_src, tile_coord(0,0,0), adj_ret); }
+template <typename T, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_load(array_t<T>& src, int x, int y, int z, int w, array_t<T>& adj_src, int adj_x, int adj_y, int adj_z, int adj_w, AdjTile& adj_ret) { adj_tile_load( src, tile_coord(x, y, z, w), adj_src, tile_coord(0,0,0,0), adj_ret); }
+
+
+
+template <typename T, typename Tile, typename AdjTile, typename Coord>
+inline CUDA_CALLABLE void adj_tile_store(array_t<T>& dest, Coord c, Tile& t, array_t<T>& adj_dest, Coord adj_c, AdjTile& adj_t)
+{  
+    tile_global_t<T, typename AdjTile::Layout::Shape> src(dest, c);
+    
+    // we allow users to override grad of src
+    if (adj_dest.data)
+        src.data.grad = adj_dest.data;        
+
+    if (src.data.grad == nullptr)
+        return;
+
+    adj_t.grad_add(src);
+}
+
+template <typename T, typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_store(array_t<T>& dest, int x, Tile& t, array_t<T>& adj_dest, int adj_x, AdjTile& adj_t) { adj_tile_store(dest, tile_coord(x), t, adj_dest, tile_coord(0), adj_t); }
+template <typename T, typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_store(array_t<T>& dest, int x, int y, Tile& t, array_t<T>& adj_dest, int adj_x, int adj_y, AdjTile& adj_t) { adj_tile_store(dest, tile_coord(x, y), t, adj_dest, tile_coord(0,0), adj_t); }
+template <typename T, typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_store(array_t<T>& dest, int x, int y, int z, Tile& t, array_t<T>& adj_dest, int adj_x, int adj_y, int adj_z, AdjTile& adj_t) { adj_tile_store(dest, tile_coord(x, y, z), t, adj_dest, tile_coord(0,0,0), adj_t); }
+template <typename T, typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_store(array_t<T>& dest, int x, int y, int z, int w, Tile& t, array_t<T>& adj_dest, int adj_x, int adj_y, int adj_z, int adj_w, AdjTile& adj_t) { adj_tile_store(dest, tile_coord(x, y, z, w), t, adj_dest, tile_coord(0,0,0,0), adj_t); }
+
+
+
+// adj_tile_atomic_add is an alias for adj_tile_store
+template <typename T, typename Tile, typename AdjTile, typename AdjRet>
+inline CUDA_CALLABLE void adj_tile_atomic_add(array_t<T>& dest, int x, Tile& t, array_t<T>& adj_dest, int adj_x, AdjTile& adj_t, AdjRet& adj_ret) { adj_tile_store(dest, tile_coord(x), t, adj_dest, tile_coord(adj_x), adj_t); }
+template <typename T, typename Tile, typename AdjTile, typename AdjRet>
+inline CUDA_CALLABLE void adj_tile_atomic_add(array_t<T>& dest, int x, int y, Tile& t, array_t<T>& adj_dest, int adj_x, int adj_y, AdjTile& adj_t, AdjRet& adj_ret) { adj_tile_store(dest, tile_coord(x, y), t, adj_dest, tile_coord(adj_x, adj_y), adj_t); }
+template <typename T, typename Tile, typename AdjTile, typename AdjRet>
+inline CUDA_CALLABLE void adj_tile_atomic_add(array_t<T>& dest, int x, int y, int z, Tile& t, array_t<T>& adj_dest, int adj_x, int adj_y, int adj_z, AdjTile& adj_t, AdjRet& adj_ret) { adj_tile_store(dest, tile_coord(x, y, z), t, adj_dest, tile_coord(adj_x, adj_y, adj_z), adj_t); }
+template <typename T, typename Tile, typename AdjTile, typename AdjRet>
+inline CUDA_CALLABLE void adj_tile_atomic_add(array_t<T>& dest, int x, int y, int z, int w, Tile& t, array_t<T>& adj_dest, int adj_x, int adj_y, int adj_z, int adj_w, AdjTile& adj_t, AdjRet& adj_ret) { adj_tile_store(dest, tile_coord(x, y, z, w), t, adj_dest, tile_coord(adj_x, adj_y, adj_z, adj_w), adj_t); }
+
+
+// unary map
+template <typename Tile, typename Fwd>
+inline CUDA_CALLABLE auto tile_map(Fwd op,
+                                   Tile &a)
+{
+    auto out = tile_register_like<Tile>();
+    auto a_reg = a.copy_to_register();
+
+    using Layout = typename decltype(out)::Layout;
+    
+    WP_PRAGMA_UNROLL
+    for (int i=0; i < Layout::NumRegs; ++i)
+    {
+        out.data[i] = op(a_reg.data[i]);
+    }
+
+    return out;
+}
+
+
+template <typename Tile, typename AdjTile, typename Fwd, typename Adj>
+inline CUDA_CALLABLE void adj_tile_map(Fwd op,
+                                       Tile& a,
+                                       Adj adj_op,
+                                       Tile& adj_a,
+                                       AdjTile& adj_ret)
+{
+    auto a_reg = a.copy_to_register();   
+    auto adj_a_reg = tile_register_like<Tile>();
+    auto adj_ret_reg = adj_ret.grad_to_register();
+
+    using Layout = typename decltype(a_reg)::Layout;
+
+    WP_PRAGMA_UNROLL
+    for (int i=0; i < Layout::NumRegs; ++i)
+    {        
+        adj_op(a_reg.data[i], adj_a_reg.data[i], adj_ret_reg.data[i]);
+    }
+
+    // write adjoints back
+    adj_a.grad_add(adj_a_reg);
+}
+
+// binary map
+template <typename TileA, typename TileB, typename Fwd>
+inline CUDA_CALLABLE auto tile_map(Fwd op,
+                                   TileA& a,
+                                   TileB& b)
+{
+    auto out = tile_register_like<TileA>();
+
+    auto a_reg = a.copy_to_register();
+    auto b_reg = b.copy_to_register();
+
+    using Layout = typename decltype(out)::Layout;
+
+    WP_PRAGMA_UNROLL
+    for (int i=0; i < Layout::NumRegs; ++i)
+    {
+        out.data[i] = op(a_reg.data[i], b_reg.data[i]);
+    }
+
+    return out;
+}
+
+
+template <typename TileA, typename TileB, typename Fwd, typename Adj, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_map(Fwd op,
+                                       TileA &a,
+                                       TileB &b,
+                                       Adj adj_op,
+                                       TileA &adj_a,
+                                       TileB &adj_b,
+                                       AdjTile &adj_ret)
+{
+    auto a_reg = a.copy_to_register();   
+    auto b_reg = b.copy_to_register();
+
+    // allocate storage for adjoints
+    auto adj_a_reg = tile_register_like<TileA>();
+    auto adj_b_reg = tile_register_like<TileB>();
+
+    auto adj_ret_reg = adj_ret.grad_to_register();
+
+    using Layout = typename decltype(a_reg)::Layout;
+
+    WP_PRAGMA_UNROLL
+    for (int i=0; i < Layout::NumRegs; ++i)
+    {
+        adj_op(a_reg.data[i], b_reg.data[i], adj_a_reg.data[i], adj_b_reg.data[i], adj_ret_reg.data[i]);
+    }
+
+    adj_a.grad_add(adj_a_reg);
+    adj_b.grad_add(adj_b_reg);
+}
+
+// wrap the operator in a lambda so that we don't have to do overload resolution for things like e.g.: wp.sin()
+// this is important because many of the builtin operators don't follow particular conventions on references for 
+// the `adj_ret` parameter, which means it's not possible to figure out the overload we need using simple casting
+#define tile_unary_map(op, a) tile_map([](auto x) { return op(x);}, a)
+#define adj_tile_unary_map(op, a, adj_op, adj_a, adj_ret) adj_tile_map([](auto x) { return op(x);}, a, [](auto x, auto& adj_x, auto adj_ret) { adj_op(x, adj_x, adj_ret);}, adj_a, adj_ret)
+
+#define tile_binary_map(op, a, b) tile_map([](auto x, auto y) { return op(x, y);}, a, b)
+#define adj_tile_binary_map(op, a, b, adj_op, adj_a, adj_b, adj_ret) adj_tile_map([](auto x, auto y) { return op(x, y);}, a, b, [](auto x, auto y, auto& adj_x, auto& adj_y, auto adj_ret) { adj_op(x, y, adj_x, adj_y, adj_ret);}, adj_a, adj_b, adj_ret)
+
+// -tile (unary neg)
+template <typename Tile>
+inline CUDA_CALLABLE auto tile_neg(Tile& a) { return tile_unary_map(wp::neg, a); }
+
+template <typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_neg(Tile& a, Tile& adj_a, AdjTile& adj_ret) { adj_tile_unary_map(wp::neg, a, wp::adj_neg, adj_a, adj_ret); }
+
+
+// tile + tile
+template <typename TileA, typename TileB>
+inline CUDA_CALLABLE auto tile_add(TileA& a, TileB& b)
+{
+    return tile_binary_map(add, a, b);
+}
+
+template <typename TileA, typename TileB, typename AdjTileA, typename AdjTileB, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_add(TileA& a, TileB& b, AdjTileA& adj_a, AdjTileB& adj_b, AdjTile& adj_c)
+{   
+    adj_tile_binary_map(add, a, b, adj_add, adj_a, adj_b, adj_c);
+}
+
+// tile - tile
+template <typename TileA, typename TileB>
+inline CUDA_CALLABLE auto tile_sub(TileA& a, TileB& b)
+{
+    return tile_binary_map(sub, a, b);
+}
+
+template <typename TileA, typename TileB, typename AdjTileA, typename AdjTileB, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_sub(TileA& a, TileB& b, AdjTileA& adj_a, AdjTileB& adj_b, AdjTile& adj_c)
+{   
+    adj_tile_binary_map(sub, a, b, adj_sub, adj_a, adj_b, adj_c);
+}
+
+
+// tile*scalar
+template <typename Tile>
+inline CUDA_CALLABLE auto tile_mul(Tile& a, const typename Tile::Type& s)
+{
+    // promote scalar to a constant tile
+    auto s_tile = tile_register_t<typename Tile::Type, tile_layout_register_t<typename Tile::Layout::Shape>>(s);
+
+    return tile_binary_map(mul, a, s_tile);
+}
+
+template <typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_mul(Tile& a, const typename Tile::Type& s,
+                                       Tile& adj_a, typename Tile::Type& adj_s,
+                                       AdjTile& adj_c)
+{
+    auto s_tile = tile_register_like<Tile>();
+    auto adj_s_tile = tile_register_like<Tile>();
+
+    using Layout = typename decltype(adj_s_tile)::Layout;
+
+    // initialize to constant
+    s_tile = s;
+
+    adj_tile_binary_map(mul, a, s_tile, adj_mul, adj_a, adj_s_tile, adj_c);
+
+    for (int i=0; i < Layout::NumRegs; ++i)
+    {
+        adj_s += adj_s_tile.data[i];
+    }
+}
+
+
+// scalar*tile
+template <typename Tile>
+inline CUDA_CALLABLE auto tile_mul(const typename Tile::Type& s, Tile& a)
+{
+    return tile_mul(a, s);
+}
+
+template <typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_mul(const typename Tile::Type& s, Tile& a,
+                                       typename Tile::Type& adj_s, Tile& adj_a,
+                                       AdjTile& adj_c)
+{
+    adj_tile_mul(a, s, adj_a, adj_s, adj_c);
+}
+
+
+template<typename Tile>
+typename Tile::Type tile_extract(Tile& t, int i) { return t.extract(tile_coord(i)); }
+template<typename Tile>
+typename Tile::Type tile_extract(Tile& t, int i, int j) { return t.extract(tile_coord(i,j)); }
+template<typename Tile>
+typename Tile::Type tile_extract(Tile& t, int i, int j, int k) { return t.extract(tile_coord(i,j,k)); }
+template<typename Tile>
+typename Tile::Type tile_extract(Tile& t, int i, int j, int k, int l) { return t.extract(tile_coord(i,j,k,l)); }
+
+
+template<typename Tile, typename AdjTile>
+void adj_tile_extract(Tile& t, int i, AdjTile& adj_t, int adj_i, typename Tile::Type adj_ret) { adj_t.adj_extract(tile_coord(i), adj_ret); }
+template<typename Tile, typename AdjTile>
+void adj_tile_extract(Tile& t, int i, int j, AdjTile& adj_t, int adj_i, int adj_j, typename Tile::Type adj_ret) { adj_t.adj_extract(tile_coord(i, j), adj_ret); }
+template<typename Tile, typename AdjTile>
+void adj_tile_extract(Tile& t, int i, int j, int k, AdjTile& adj_t, int adj_i, int adj_j, int adj_k, typename Tile::Type adj_ret) { adj_t.adj_extract(tile_coord(i, j, k), adj_ret); }
+template<typename Tile, typename AdjTile>
+void adj_tile_extract(Tile& t, int i, int j, int k, int l, AdjTile& adj_t, int adj_i, int adj_j, int adj_k, int adj_l, typename Tile::Type adj_ret) { adj_t.adj_extract(tile_coord(i, j, k, l), adj_ret); }
+
+
+namespace partitioned_gemm
+{
+
+template <typename T>
+inline CUDA_CALLABLE const T& index(const T* __restrict__ p, int i, int j, int stride)
+{
+    return p[i*stride + j];
+}
+
+template <typename T>
+inline CUDA_CALLABLE T& index(T* __restrict__ p, int i, int j, int stride)
+{
+    return p[i*stride + j];
+}
+
+template <int PartitionM, int PartitionN, typename Tile>
+struct partition_t
+{
+    static constexpr int M = PartitionM;
+    static constexpr int N = PartitionN;
+    static constexpr int Stride = Tile::Layout::Shape::dim(1);
+    
+    using T = typename Tile::Type;    
+
+    inline partition_t(Tile& A) 
+    {
+        data = A.data.ptr;
+        
+        // todo: do ceil div for non-multiples of M,N
+        shape[0] = Tile::Layout::Shape::dim(0)/PartitionM;
+        shape[1] = Tile::Layout::Shape::dim(1)/PartitionN;
+    }
+
+    // underlying data
+    T* data;
+    
+    // partition dimensions
+    int shape[2];
+};
+
+template <typename Partition>
+inline int partition_size(const Partition& part)
+{
+    return part.shape[0]*part.shape[1];
+}
+
+// returns the x, y coordinates of a tile given a linear index
+template <typename Partition>
+inline void partition_coord(const Partition& part, const int t, int& i, int& j)
+{
+    i = t/part.shape[1];
+    j = t%part.shape[1];
+}
+
+template <typename Partition>
+inline auto partition_load(const Partition& tile, int i, int j)
+{
+    mat_t<Partition::M, Partition::N, typename Partition::T> out;
+    
+    const int tile_i = i*Partition::M;
+    const int tile_j = j*Partition::N;
+
+    WP_PRAGMA_UNROLL
+    for (int i=0; i < Partition::M; ++i)
+    {
+        WP_PRAGMA_UNROLL
+        for (int j=0; j < Partition::N; ++j)
+        {
+            out.data[i][j] = partitioned_gemm::index(tile.data, tile_i + i, tile_j + j, Partition::Stride);
+        }
+    }
+
+    return out;
+}
+
+template <typename Partition, typename Value>
+inline void partition_store(const Partition& tile, int i, int j, const Value& value)
+{
+    const int tile_i = Partition::M*i;
+    const int tile_j = Partition::N*j;
+
+    WP_PRAGMA_UNROLL
+    for (int i=0; i < Partition::M; ++i)
+    {	
+        WP_PRAGMA_UNROLL
+        for (int j=0; j < Partition::N; ++j)
+        {
+            index(tile.data, tile_i + i, tile_j + j, Partition::Stride) = value.data[i][j];
+        }
+    }
+}
+
+
+template <typename TileA, typename TileB, typename TileC>
+inline CUDA_CALLABLE void matmul(TileA& A, TileB& B, TileC& out)
+{   
+    const int TILE_M = 4;
+    const int TILE_N = 4;
+    const int TILE_K = 4;
+
+    auto A_tile = partition_t<TILE_M, TILE_K, TileA>(A);
+    auto B_tile = partition_t<TILE_K, TILE_N, TileB>(B);
+    auto C_tile = partition_t<TILE_M, TILE_N, TileC>(out);
+
+    //static_assert(is_same<typename TileA::Type, typename TileB::Type>::value);
+
+    const int length = partition_size(C_tile);
+
+    for (int t=WP_TILE_THREAD_IDX; t < length; t += WP_TILE_BLOCK_DIM)
+    {  
+        int i, j;
+        partition_coord(C_tile, t, i, j);
+
+        // accumulator
+        auto sum = partition_load(C_tile, i, j);
+
+        WP_PRAGMA_UNROLL
+        for (int k=0; k < A_tile.shape[1]; k++)
+        {
+            const auto a = partition_load(A_tile, i, k);
+            const auto b = partition_load(B_tile, k, j);
+
+            sum += mul(a, b);
+        }
+        
+        partition_store(C_tile, i, j, sum);
+    }
+}
+
+template <typename LayoutA, typename LayoutB, typename LayoutC, typename StorageA, typename StorageB, typename StorageC, typename T>
+inline CUDA_CALLABLE void scalar_matmul(const StorageA& A, const StorageB& B, StorageC& C, T scale)
+{
+    for (int t=WP_TILE_THREAD_IDX; t < LayoutC::Size; t += WP_TILE_BLOCK_DIM)
+    {  
+        auto coord = LayoutC::coord_from_linear(t);
+
+        int i = coord[0];
+        int j = coord[1];
+
+        // accumulator
+        auto sum = C(coord)*scale;
+
+        WP_PRAGMA_UNROLL
+        for (int k=0; k < LayoutA::Shape::dim(1); k++)
+        {
+            const auto a = A(tile_coord(i, k));
+            const auto b = B(tile_coord(k, j));
+
+            sum = muladd<decltype(sum)>(a, b, sum);
+        }
+        
+        C(coord) = sum;
+    }
+}
+
+template <typename TileA, typename TileL>
+inline CUDA_CALLABLE void scalar_cholesky(TileA& A, TileL& L)
+{
+    using T = typename TileA::Type;    
+    constexpr int n = TileA::Layout::Shape::dim(1);
+
+    for (int j=0; j < n; ++j)
+    {
+        T s = A.data(tile_coord(j, j));
+
+        for (int k=0; k < j; ++k) 
+        {
+            T r = L.data(tile_coord(j, k));
+            s -= r * r;
+        }
+
+        s = wp::sqrt(s);
+        T invS = 1.0 / s;
+
+        L.data(tile_coord(j, j)) = s;
+
+        for (int i=j+1; i < n; ++i)
+        {
+            s = A.data(tile_coord(i, j));
+
+            for (int k=0; k < j; ++k)
+            {
+                s -= L.data(tile_coord(i, k)) * L.data(tile_coord(j, k));
+            }
+
+            L.data(tile_coord(i, j)) = s * invS;           
+        }
+
+        // zero out upper triangular portion
+        for (int k=j+1; k < n; ++k)
+        {
+            L.data(tile_coord(j,k)) = T(0.0);
+        }
+    }
+}
+
+template <typename TileL, typename TileX, typename TileY>
+inline CUDA_CALLABLE void scalar_cholesky_solve(TileL& L, TileX& X, TileY& Y)
+{
+    using T = typename TileL::Type;    
+    constexpr int n = TileL::Layout::Shape::dim(1);
+
+    for (int i=0; i < n; ++i)
+    {
+        T s = Y.data(tile_coord(i));
+
+        for (int j=0; j < i; ++j)
+            s -= L.data(tile_coord(i,j)) * X.data(tile_coord(j));
+
+        X.data(tile_coord(i)) = s / L.data(tile_coord(i, i));
+    }
+
+    for (int i=n-1; i >= 0; --i)
+    {
+        T s = X.data(tile_coord(i));
+
+        for (int j=i+1; j < n; ++j)
+            s -= L.data(tile_coord(j, i)) * X.data(tile_coord(j));
+
+        X.data(tile_coord(i)) = s / L.data(tile_coord(i, i));
+    }
+}
+
+} // namespace partition_gemm
+
+
+template <int Add, typename Fwd, typename AdjA, typename AdjB, typename TileA, typename TileB, typename TileC>
+TileC& tile_matmul(Fwd fun_forward, AdjA fun_backward_A, AdjB fun_backward_B, TileA& A, TileB& B, TileC& C)
+{       
+    using ShapeA = typename TileA::Layout::Shape;
+    using ShapeB = typename TileB::Layout::Shape;
+    using ShapeC = typename TileC::Layout::Shape;
+
+    static_assert(ShapeA::N == 2, "Expected ShapeA::N == 2");
+    static_assert(ShapeB::N == 2, "Expected ShapeB::N == 2");
+    static_assert(ShapeC::N == 2, "Expected ShapeC::N == 2");
+
+    static_assert(ShapeA::dim(1) == ShapeB::dim(0), "Expected ShapeA::dim(1) == ShapeB::dim(0)");
+    static_assert(ShapeC::dim(0) == ShapeA::dim(0), "Expected ShapeC::dim(0) == ShapeA::dim(0)");
+    static_assert(ShapeC::dim(1) == ShapeB::dim(1), "Expected ShapeC::dim(1) == ShapeB::dim(1)");
+    
+
+    using T = typename TileA::Type;
+
+#if !defined(__CUDA_ARCH__) || WP_ENABLE_MATHDX == 0
+    partitioned_gemm::scalar_matmul<typename TileA::Layout, typename TileB::Layout, typename TileC::Layout>(A.data, B.data, C.data, T(Add));
+#else
+    fun_forward(T(1.0), A.data.ptr, B.data.ptr, T(Add), C.data.ptr);
+#endif
+    
+    WP_TILE_SYNC();
+    
+    return C;
+}
+
+
+// backward for the wp.tile_matmul(a, b, out) syntax
+template <typename Fwd, typename AdjA, typename AdjB, typename TileA, typename TileB, typename TileC>
+void adj_tile_matmul(Fwd fun_forward, AdjA fun_backward_A, AdjB fun_backward_B, TileA& A, TileB& B, TileC& C,
+                   Fwd adj_fun_forward, AdjA adj_fun_backward_A, AdjB adj_fun_backward_B, TileA& adj_A, TileB& adj_B, TileC& adj_C)
+{   
+    using T = typename TileA::Type;    
+
+#if !defined(__CUDA_ARCH__) || WP_ENABLE_MATHDX == 0
+    auto At = tile_transpose(A);
+    auto Bt = tile_transpose(B);
+
+    partitioned_gemm::scalar_matmul<typename TileC::Layout, typename decltype(Bt)::Layout, typename TileA::Layout>(adj_C.grad, Bt.data, adj_A.grad, T(1.0));
+    partitioned_gemm::scalar_matmul<typename decltype(At)::Layout, typename TileC::Layout, typename TileB::Layout>(At.data, adj_C.grad, adj_B.grad, T(1.0));
+#else
+    fun_backward_A(T(1.0), adj_C.grad.ptr, B.data.ptr, T(1.0), adj_A.grad.ptr);
+    fun_backward_B(T(1.0), A.data.ptr, adj_C.grad.ptr, T(1.0), adj_B.grad.ptr);
+#endif
+
+    WP_TILE_SYNC();
+}
+
+// backward for the out = wp.tile_matmul(a, b) syntax
+template <typename Fwd, typename AdjA, typename AdjB, typename TileA, typename TileB, typename TileC>
+void adj_tile_matmul(Fwd fun_forward, AdjA fun_backward_A, AdjB fun_backward_B, TileA& A, TileB& B, TileC& C, 
+                   Fwd adj_fun_forward, AdjA adj_fun_backward_A, AdjB adj_fun_backward_B, TileA& adj_A, TileB& adj_B, TileC& adj_C, TileC& adj_ret)
+{   
+    using T = typename TileA::Type;    
+
+#if !defined(__CUDA_ARCH__) || WP_ENABLE_MATHDX == 0
+    auto At = tile_transpose(A);
+    auto Bt = tile_transpose(B);
+
+    partitioned_gemm::scalar_matmul<typename TileC::Layout, typename decltype(Bt)::Layout, typename TileA::Layout>(adj_C.grad, Bt.data, adj_A.grad, T(1.0));
+    partitioned_gemm::scalar_matmul<typename decltype(At)::Layout, typename TileC::Layout, typename TileB::Layout>(At.data, adj_C.grad, adj_B.grad, T(1.0));
+#else
+    fun_backward_A(T(1.0), adj_C.grad.ptr, B.data.ptr, T(1.0), adj_A.grad.ptr);
+    fun_backward_B(T(1.0), A.data.ptr, adj_C.grad.ptr, T(1.0), adj_B.grad.ptr);
+#endif
+
+    WP_TILE_SYNC();
+}
+
+#if !defined(__CUDA_ARCH__) || WP_ENABLE_MATHDX == 0
+
+#define tile_fft()
+#define tile_ifft()
+
+#define adj_tile_fft()
+#define adj_tile_ifft()
+
+#else
+
+// TODO(lcambier): use a properly overaligned complex type that matches cuFFTDx's expectation
+// and remove the need for __align__(16) dtypes data[...]
+#define tile_fft(function_name, dtype, shared_memory_size, batch_size, ept, Xinout) \
+    do { \
+        void function_name(dtype*, dtype*); \
+        char* buffer = (char*)wp::tile_alloc_shared(shared_memory_size); \
+        __align__(16) dtype data[ept]; \
+        for(int b = 0; b < (int)batch_size; b++) { \
+            dtype* inout = Xinout.data + (int)b * (int)ept; \
+            memcpy(data, inout, sizeof(dtype) * ept); \
+            function_name(data, (dtype*)buffer); \
+            memcpy(inout, data, sizeof(dtype) * ept); \
+            WP_TILE_SYNC(); \
+        } \
+        wp::tile_alloc_shared(-shared_memory_size); \
+    } while (0)
+
+#define tile_ifft tile_fft
+
+// adj_function_name, adj_dtype, adj_shared_memory_size, adj_batch_size, adj_ept are all ignored
+
+#define adj_tile_fft(function_name, dtype, shared_memory_size, batch_size, ept, Xinout, \
+                        adj_function_name, adj_dtype, adj_shared_memory_size, adj_batch_size, adj_ept, \
+                        adj_Xinout) \
+    do { \
+        tile_ifft(function_name, dtype, shared_memory_size, batch_size, ept, adj_Xinout); \
+    } while (0)
+
+#define adj_tile_ifft(function_name, dtype, shared_memory_size, batch_size, ept, Xinout, \
+                         adj_function_name, adj_dtype, adj_shared_memory_size, adj_batch_size, adj_ept, \
+                         adj_Xinout) \
+    do { \
+        tile_fft(function_name, dtype, shared_memory_size, batch_size, ept, adj_Xinout); \
+    } while (0)
+
+#endif // !defined(__CUDA_ARCH__)
+
+template <typename Fwd, typename TileA, typename TileL>
+TileL& tile_cholesky(Fwd fun_forward, TileA& A, TileL& L)
+{       
+    // Copy to L
+    L = A;
+
+#if !defined(__CUDA_ARCH__) || WP_ENABLE_MATHDX == 0
+
+    partitioned_gemm::scalar_cholesky(A, L);
+
+#else
+
+
+    // Call cholesky on L
+    WP_TILE_SYNC();
+    
+    fun_forward(L.data.ptr, TileL::Layout::Shape::dim(0));
+    
+    WP_TILE_SYNC();
+
+    // Zero-out the upper triangular part of L
+
+    WP_PRAGMA_UNROLL
+    for (int i=WP_TILE_THREAD_IDX; i < TileL::Layout::Size; i += WP_TILE_BLOCK_DIM)
+    {
+        auto c = TileL::Layout::coord_from_linear(i);
+        
+        if(c[0] < c[1]) 
+            L.data(c) = 0.0;
+    }
+
+    WP_TILE_SYNC();
+
+#endif
+
+    return L;
+}
+
+#define adj_tile_cholesky(function_name, A, L, \
+                          adj_function_name, adj_A, adj_L, adj_ret) \
+    do { \
+        assert(false); \
+    } while (0)
+
+template <typename Fwd, typename TileL, typename TileX, typename TileY>
+TileY& tile_cholesky_solve(Fwd fun_forward, TileL& L, TileX& X, TileY& Y)
+{       
+    // Copy x to y
+
+    Y = X;
+
+#if !defined(__CUDA_ARCH__) || WP_ENABLE_MATHDX == 0
+
+    partitioned_gemm::scalar_cholesky_solve(L, X, Y);
+
+#else
+
+    // Call cholesky solve on L & y
+
+    WP_TILE_SYNC();
+    
+    fun_forward(L.data.ptr, Y.data.ptr); \
+
+    WP_TILE_SYNC();
+    
+#endif
+
+    return Y;
+}
+
+#define adj_tile_cholesky_solve(function_name, L, X, Y, \
+                                adj_function_name, adj_L, adj_X, adj_Y, adj_ret) \
+    do { \
+        assert(false); \
+    } while (0)
+
+template <typename Tile>
+inline CUDA_CALLABLE auto tile_transpose(Tile& t)
+{    
+    static_assert(Tile::Layout::Shape::N == 2, "Expected Tile::Layout::Shape::N == 2");
+
+    // alias incoming tile 
+    constexpr int M = Tile::Layout::Shape::dim(0);
+    constexpr int N = Tile::Layout::Shape::dim(1);
+
+    constexpr int StrideM = Tile::Layout::Stride::dim(0);
+    constexpr int StrideN = Tile::Layout::Stride::dim(1);
+
+    return tile_shared_t<typename Tile::Type, tile_layout_strided_t<tile_shape_t<N,M>, tile_stride_t<StrideN, StrideM>>, false>(t.data.ptr, t.grad.ptr);
+}
+
+template <typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_transpose(Tile& t, Tile& adj_t, AdjTile& adj_ret)
+{    
+    auto a = tile_transpose(adj_ret);
+    auto b = adj_t;
+    
+    adj_t.assign(tile_add(a,b));
+}
+
+template <int N, int StrideN, typename Tile>
+inline CUDA_CALLABLE auto tile_broadcast(Tile& t)
+{
+    // alias incoming tile with new strides
+    return tile_shared_t<typename Tile::Type, tile_layout_strided_t<tile_shape_t<N>, tile_stride_t<StrideN>>, false>(t.data.ptr, t.grad.ptr);
+}
+
+template <int M, int N, int StrideM, int StrideN, typename Tile>
+inline CUDA_CALLABLE auto tile_broadcast(Tile& t)
+{
+    // alias incoming tile with new strides
+    return tile_shared_t<typename Tile::Type, tile_layout_strided_t<tile_shape_t<M, N>, tile_stride_t<StrideM, StrideN>>, false>(t.data.ptr, t.grad.ptr);
+}
+
+template <int M, int N, int O, int StrideM, int StrideN, int StrideO, typename Tile>
+inline CUDA_CALLABLE auto tile_broadcast(Tile& t)
+{
+    // alias incoming tile with new strides
+    return tile_shared_t<typename Tile::Type, tile_layout_strided_t<tile_shape_t<M, N, O>, tile_stride_t<StrideM, StrideN, StrideO>>, false>(t.data.ptr, t.grad.ptr);
+}
+
+template <int M, int N, int O, int P, int StrideM, int StrideN, int StrideO, int StrideP, typename Tile>
+inline CUDA_CALLABLE auto tile_broadcast(Tile& t)
+{
+    // alias incoming tile with new strides
+    return tile_shared_t<typename Tile::Type, tile_layout_strided_t<tile_shape_t<M, N, O, P>, tile_stride_t<StrideM, StrideN, StrideO, StrideP>>, false>(t.data.ptr, t.grad.ptr);
+}
+
+template <typename Tile, typename AdjTile>
+inline CUDA_CALLABLE void adj_tile_broadcast(Tile& t, Tile& adj_t, AdjTile& adj_ret)
+{   
+    // nop, since memory is aliased grads already accumulated
+}
+
+template <typename ReturnType, typename Tile, typename... Indices>
+inline CUDA_CALLABLE auto tile_view(Tile& t, Indices... indices)
+{   
+    auto c = tile_coord(indices...);
+
+    // return new tile with same strides
+    typename Tile::Type* data_ptr = &t.data(c);
+    typename Tile::Type* grad_ptr = nullptr;
+    
+    if (t.grad.ptr)
+        grad_ptr = &t.grad(c);
+
+    return ReturnType(data_ptr, grad_ptr);
+}
+
+
+template <typename TileA, typename Scalar>
+inline CUDA_CALLABLE void assign(TileA& dest, int i, const Scalar& src)
+{   
+    dest.data(tile_coord(i)) = src;
+    WP_TILE_SYNC();
+}
+
+template <typename TileA, typename Scalar>
+inline CUDA_CALLABLE void assign(TileA& dest, int i, int j, const Scalar& src)
+{   
+    dest.data(tile_coord(i, j)) = src;
+    WP_TILE_SYNC();
+}
+
+template <typename TileA, typename Scalar>
+inline CUDA_CALLABLE void assign(TileA& dest, int i, int j, int k, const Scalar& src)
+{   
+    dest.data(tile_coord(i, j, k)) = src;
+    WP_TILE_SYNC();
+}
+
+template <typename TileA, typename Scalar>
+inline CUDA_CALLABLE void assign(TileA& dest, int i, int j, int k, int l, const Scalar& src)
+{   
+    dest.data(tile_coord(i, j, k, l)) = src;
+    WP_TILE_SYNC();
+}
+
+
+
+
+template <typename TileA, typename TileB, typename Coord>
+inline CUDA_CALLABLE void tile_assign(TileA& dest, TileB& src, const Coord& offset)
+{   
+    using Layout = typename TileB::Layout;
+
+    for (int t=WP_TILE_THREAD_IDX; t < Layout::Size; t += WP_TILE_BLOCK_DIM)
+    {
+        auto c = Layout::coord_from_linear(t);
+        dest.data(c + offset) = src.data(c);
+    }
+
+    WP_TILE_SYNC();
+}
+
+template <typename TileA, typename TileB, typename AdjTileA, typename AdjTileB, typename Coord, typename AdjCoord>
+inline CUDA_CALLABLE void adj_tile_assign(TileA& dest, TileB& src, Coord offset,
+                                          AdjTileA& adj_dest, AdjTileB& adj_src, AdjCoord adj_offset)
+{
+    using Layout = typename TileB::Layout;
+
+    for (int t=WP_TILE_THREAD_IDX; t < Layout::Size; t += WP_TILE_BLOCK_DIM)
+    {
+        auto c = Layout::coord_from_linear(t);        
+        src.grad(c) += dest.grad(c + offset);
+    } 
+
+    WP_TILE_SYNC();
+}
+
+
+// codegen entry points, which emit calls like `tile_assign(dest, src, i, j, k)`
+// a better approach here would be for codegen to just directly generate `tile_assign(dest, src, tile_coord(i, j, k))` 
+// i.e.: call the above implementation methods directly, then we could remove these overloads
+template <typename TileA, typename TileB>
+inline CUDA_CALLABLE void tile_assign(TileA& dest, TileB& src, int i) { tile_assign(dest, src, tile_coord(i)); }
+template <typename TileA, typename TileB>
+inline CUDA_CALLABLE void tile_assign(TileA& dest, TileB& src, int i, int j) { tile_assign(dest, src, tile_coord(i, j)); }
+template <typename TileA, typename TileB>
+inline CUDA_CALLABLE void tile_assign(TileA& dest, TileB& src, int i, int j, int k) { tile_assign(dest, src, tile_coord(i, j, k)); }
+template <typename TileA, typename TileB>
+inline CUDA_CALLABLE void tile_assign(TileA& dest, TileB& src, int i, int j, int k, int l) { tile_assign(dest, src, tile_coord(i, j, k, l)); }
+
+template <typename TileA, typename TileB, typename AdjTileA, typename AdjTileB>
+inline CUDA_CALLABLE void adj_tile_assign(TileA& dest, TileB& src, int i, AdjTileA& adj_dest, AdjTileB& adj_src, int) { adj_tile_assign(dest, src, tile_coord(i), adj_dest, adj_src, tile_coord(0)); }
+template <typename TileA, typename TileB, typename AdjTileA, typename AdjTileB>
+inline CUDA_CALLABLE void adj_tile_assign(TileA& dest, TileB& src, int i, int j, AdjTileA& adj_dest, AdjTileB& adj_src, int, int) { adj_tile_assign(dest, src, tile_coord(i,j), adj_dest, adj_src, tile_coord(0)); }
+template <typename TileA, typename TileB, typename AdjTileA, typename AdjTileB>
+inline CUDA_CALLABLE void adj_tile_assign(TileA& dest, TileB& src, int i, int j, int k, AdjTileA& adj_dest, AdjTileB& adj_src, int, int, int) { adj_tile_assign(dest, src, tile_coord(i,j,k), adj_dest, adj_src, tile_coord(0)); }
+template <typename TileA, typename TileB, typename AdjTileA, typename AdjTileB>
+inline CUDA_CALLABLE void adj_tile_assign(TileA& dest, TileB& src, int i, int j, int k, int l, AdjTileA& adj_dest, AdjTileB& adj_src, int, int, int, int) { adj_tile_assign(dest, src, tile_coord(i,j,k,l), adj_dest, adj_src, tile_coord(0)); }
+
+
+template <typename TileA, typename TileB, typename TileC>
+inline CUDA_CALLABLE TileC& tile_diag_add(TileA& a, TileB& b, TileC& c)
+{   
+    using ShapeA = typename TileA::Layout::Shape;
+    using ShapeB = typename TileB::Layout::Shape;
+    using ShapeC = typename TileC::Layout::Shape;
+
+    static_assert(ShapeA::dim(0) == ShapeA::dim(1), "Expected ShapeA::dim(0) == ShapeA::dim(1)");
+    static_assert(ShapeB::dim(0) == ShapeA::dim(0), "Expected ShapeB::dim(0) == ShapeA::dim(0)");
+    static_assert(ShapeC::dim(0) == ShapeA::dim(0), "Expected ShapeC::dim(0) == ShapeA::dim(0)");
+    static_assert(ShapeC::dim(0) == ShapeC::dim(1), "Expected ShapeC::dim(0) == ShapeC::dim(1)");
+
+    c = a;
+    
+    for (int t=WP_TILE_THREAD_IDX; t < ShapeA::dim(0); t += WP_TILE_BLOCK_DIM)
+    {
+        c.data(tile_coord(t, t)) += b.data(tile_coord(t));
+    }
+
+    WP_TILE_SYNC();
+
+    return c;
+}
+
+template <typename TileA, typename TileB, typename TileC, typename AdjTileA, typename AdjTileB, typename AdjTileC>
+inline CUDA_CALLABLE void adj_tile_diag_add(TileA& a, TileB& b, TileC& c, AdjTileA& adj_a, AdjTileB& adj_b, AdjTileC& adj_c, AdjTileC& adj_ret)
+{   
+    assert(false);
+}
+
+
+} // namespace wp
+
+
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif