Winograd Update for Small batches (#1803)

* Build in padding to Winograd kernels
* Add new fused Winograd kernel
* Enable weight flipping in Winograd kernels

mlx/backend/metal/conv.cpp

@@ -533,6 +533,45 @@ void implicit_gemm_conv_2D_general_gpu(
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
 
+void winograd_conv_2D_fused_gpu(
+    const Stream& s,
+    metal::Device& d,
+    const array& in,
+    const array& wt,
+    array out,
+    const MLXConvParams<2>& conv_params,
+    std::vector<array>& copies_w) {
+  int O_c = conv_params.O;
+  int C_c = conv_params.C;
+
+  int N_tiles_n = conv_params.N;
+  int N_tiles_h = (conv_params.oS[0] + 1) / 2;
+  int N_tiles_w = (conv_params.oS[1] + 1) / 2;
+  int N_tiles = N_tiles_n * N_tiles_h * N_tiles_w;
+
+  int bc = 32;
+  int wm = 4;
+  int wn = 1;
+  std::ostringstream kname;
+  kname << "winograd_conv_2d_fused_" << type_to_name(out) << "_flip"
+        << conv_params.flip;
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = d.get_kernel(kname.str());
+  compute_encoder.set_compute_pipeline_state(kernel);
+
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
+
+  compute_encoder.set_bytes(conv_params, 3);
+
+  MTL::Size group_dims = MTL::Size(8, 8, 2);
+  MTL::Size grid_dims =
+      MTL::Size(O_c / 8, (N_tiles_h * N_tiles_w) / 8, N_tiles_n);
+
+  compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
+}
+
 void winograd_conv_2D_gpu(
     const Stream& s,
     metal::Device& d,
@@ -541,67 +580,6 @@ void winograd_conv_2D_gpu(
     array out,
     const MLXConvParams<2>& conv_params,
     std::vector<array>& copies_w) {
-  Shape padded_shape = {
-      conv_params.N,
-      conv_params.iS[0] + 2 * conv_params.pad[0],
-      conv_params.iS[1] + 2 * conv_params.pad[1],
-      conv_params.C};
-
-  padded_shape[1] = 6 * ((padded_shape[1] - 2 + 5) / 6) + 2;
-  padded_shape[2] = 6 * ((padded_shape[2] - 2 + 5) / 6) + 2;
-
-  array in_padded(std::move(padded_shape), in.dtype(), nullptr, {});
-
-  // Fill with zeros
-  array zero_arr = array(0, in.dtype());
-  fill_gpu(zero_arr, in_padded, s);
-  copies_w.push_back(zero_arr);
-
-  // Pick input slice from padded
-  size_t data_offset = conv_params.pad[0] * in_padded.strides()[1] +
-      conv_params.pad[1] * in_padded.strides()[2];
-  array in_padded_slice(in.shape(), in_padded.dtype(), nullptr, {});
-  in_padded_slice.copy_shared_buffer(
-      in_padded,
-      in_padded.strides(),
-      in_padded.flags(),
-      in_padded_slice.size(),
-      data_offset);
-
-  // Copy input values into the slice
-  copy_gpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, s);
-
-  copies_w.push_back(in_padded_slice);
-  copies_w.push_back(in_padded);
-
-  MLXConvParams<2> conv_params_updated{
-      /* const int  N = */ static_cast<int>(in_padded.shape(0)),
-      /* const int  C = */ static_cast<int>(in_padded.shape(3)),
-      /* const int  O = */ static_cast<int>(wt.shape(0)),
-      /* const int iS[NDIM] = */
-      {static_cast<int>(in_padded.shape(1)),
-       static_cast<int>(in_padded.shape(2))},
-      /* const int wS[NDIM] = */
-      {static_cast<int>(wt.shape(1)), static_cast<int>(wt.shape(2))},
-      /* const int oS[NDIM] = */
-      {static_cast<int>(out.shape(1)), static_cast<int>(out.shape(2))},
-      /* const int str[NDIM] = */ {1, 1},
-      /* const int pad[NDIM] = */ {0, 0},
-      /* const int kdil[NDIM] = */ {1, 1},
-      /* const int idil[NDIM] = */ {1, 1},
-      /* const size_t in_strides[NDIM + 2] = */
-      {in_padded.strides()[0],
-       in_padded.strides()[1],
-       in_padded.strides()[2],
-       in_padded.strides()[3]},
-      /* const size_t wt_strides[NDIM + 2] = */
-      {wt.strides()[0], wt.strides()[1], wt.strides()[2], wt.strides()[3]},
-      /* const size_t out_strides[NDIM + 2] = */
-      {out.strides()[0], out.strides()[1], out.strides()[2], out.strides()[3]},
-      /* const int groups = */ 1,
-      /* const bool flip = */ false,
-  };
-
   int O_c = conv_params.O;
   int C_c = conv_params.C;
 
@@ -620,7 +598,7 @@ void winograd_conv_2D_gpu(
     int bo = 4;
     std::ostringstream kname;
     kname << "winograd_conv_2d_weight_transform_" << type_to_name(out) << "_bc"
-          << bc;
+          << bc << "_flip" << conv_params.flip;
     auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder.set_compute_pipeline_state(kernel);
@@ -653,10 +631,10 @@ void winograd_conv_2D_gpu(
     auto kernel = d.get_kernel(kname.str());
     compute_encoder.set_compute_pipeline_state(kernel);
 
-    compute_encoder.set_input_array(in_padded, 0);
+    compute_encoder.set_input_array(in, 0);
     compute_encoder.set_output_array(inp_wg, 1);
 
-    compute_encoder.set_bytes(conv_params_updated, 2);
+    compute_encoder.set_bytes(conv_params, 2);
 
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(N_tiles_w, N_tiles_h, N_tiles_n);
@@ -703,7 +681,7 @@ void winograd_conv_2D_gpu(
     compute_encoder.set_input_array(out_wg, 0);
     compute_encoder.set_output_array(out, 1);
 
-    compute_encoder.set_bytes(conv_params_updated, 2);
+    compute_encoder.set_bytes(conv_params, 2);
 
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(N_tiles_w, N_tiles_h, N_tiles_n);
@@ -767,15 +745,19 @@ void conv_2D_gpu(
   }
 
   // Direct to winograd conv
-  bool inp_large =
+  bool img_large =
       (conv_params.N * conv_params.iS[0] * conv_params.iS[1]) >= 1ul << 12;
   bool channels_large = (conv_params.C + conv_params.O) >= 256;
-  if (!flip && is_stride_one && is_kdil_one && is_idil_one &&
+  if (conv_params.wS[0] == 3 && conv_params.wS[1] == 3 &&
-      conv_params.wS[0] == 3 && conv_params.wS[1] == 3 &&
+      conv_params.C % 32 == 0 && conv_params.O % 32 == 0 && is_stride_one &&
-      conv_params.C % 32 == 0 && conv_params.O % 32 == 0 && inp_large &&
+      is_kdil_one && is_idil_one) {
-      channels_large) {
+    if (img_large && channels_large) {
       return winograd_conv_2D_gpu(s, d, in, wt, out, conv_params, copies);
     }
+    if (conv_params.N <= 1) {
+      return winograd_conv_2D_fused_gpu(s, d, in, wt, out, conv_params, copies);
+    }
+  }
 
   // Direct to implicit gemm conv
   if (is_idil_one && (conv_params.C <= 4 || conv_params.C % 16 == 0) &&
@@ -876,8 +858,40 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
     wt = arr_copy;
   }
 
+  // Check for 1x1 conv
+  auto is_one = [](int x) { return x == 1; };
+  auto is_zero = [](int x) { return x == 0; };
+  if (groups_ == 1 && (wt.shape(0) * wt.shape(-1) == wt.size()) &&
+      std::all_of(wt.shape().begin() + 1, wt.shape().end() - 1, is_one) &&
+      std::all_of(kernel_strides_.begin(), kernel_strides_.end(), is_one) &&
+      std::all_of(input_dilation_.begin(), input_dilation_.end(), is_one) &&
+      std::all_of(kernel_dilation_.begin(), kernel_dilation_.end(), is_one) &&
+      std::all_of(padding_.begin(), padding_.end(), is_zero)) {
+    std::vector<array> empty_copies;
+    steel_matmul_regular(
+        s,
+        d,
+        /*a = */ in,
+        /*b = */ wt,
+        /*c = */ out,
+        /*M = */ in.size() / in.shape(-1),
+        /*N = */ wt.shape(0),
+        /*K = */ in.shape(-1),
+        /*batch_size_out = */ 1,
+        /*lda = */ in.shape(-1),
+        /*ldb = */ wt.shape(-1),
+        /*ldd = */ wt.shape(0),
+        /*transpose_a = */ false,
+        /*transpose_b = */ true,
+        /*batch_shape = */ {1},
+        /*batch_strides = */ {1},
+        /*A_batch_stride = */ 0,
+        /*B_batch_stride = */ 0,
+        /*matrix_stride_out = */ 0,
+        /*copies = */ empty_copies);
+  }
   // 3D conv
-  if (out.ndim() == 5) {
+  else if (out.ndim() == 5) {
     conv_3D_gpu(
         s,
         d,

mlx/backend/metal/kernels/conv.metal

@@ -326,7 +326,13 @@ constant constexpr const float WinogradTransforms<6, 3, 8>::wt_transform[8][8];
 constant constexpr const float WinogradTransforms<6, 3, 8>::in_transform[8][8];
 constant constexpr const float WinogradTransforms<6, 3, 8>::out_transform[8][8];
 
-template <typename T, int BC = 32, int BO = 4, int M = 6, int R = 3>
+template <
+    typename T,
+    int BC = 32,
+    int BO = 4,
+    bool do_flip = false,
+    int M = 6,
+    int R = 3>
 [[kernel, max_total_threads_per_threadgroup(BO * 32)]] void
 winograd_conv_2d_weight_transform(
     const device T* wt_in [[buffer(0)]],
@@ -373,10 +379,15 @@ winograd_conv_2d_weight_transform(
     for (int kh = 0; kh < R; ++kh) {
       for (int kw = 0; kw < R; ++kw) {
         for (int kc = simd_lane_id; kc < BC; kc += 32) {
+          if (do_flip) {
+            Ws[simd_group_id][R - 1 - kh][R - 1 - kw][kc] =
+                wt_in[kh * R * C + kw * C + kc];
+          } else {
             Ws[simd_group_id][kh][kw][kc] = wt_in[kh * R * C + kw * C + kc];
           }
         }
       }
+    }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
     // Do transform and store the result
@@ -398,10 +409,10 @@ winograd_conv_2d_weight_transform(
   }
 }
 
-#define instantiate_winograd_conv_2d_weight_transform_base(name, itype, bc) \
+#define instantiate_winograd_conv_2d_weight_tr_base_2(name, itype, bc, f)   \
-  template [[host_name("winograd_conv_2d_weight_transform_" #name           \
+  template [[host_name("winograd_conv_2d_weight_transform_" #name "_bc" #bc \
-                       "_bc" #bc)]] [[kernel]] void                         \
+                       "_flip" #f)]] [[kernel]] void                        \
-  winograd_conv_2d_weight_transform<itype, bc>(                             \
+  winograd_conv_2d_weight_transform<itype, bc, 4, f>(                       \
       const device itype* wt_in [[buffer(0)]],                              \
       device itype* wt_out [[buffer(1)]],                                   \
       const constant int& C [[buffer(2)]],                                  \
@@ -410,6 +421,10 @@ winograd_conv_2d_weight_transform(
       uint simd_group_id [[simdgroup_index_in_threadgroup]],                \
       uint simd_lane_id [[thread_index_in_simdgroup]]);
 
+#define instantiate_winograd_conv_2d_weight_transform_base(name, itype, bc) \
+  instantiate_winograd_conv_2d_weight_tr_base_2(name, itype, bc, 0)         \
+      instantiate_winograd_conv_2d_weight_tr_base_2(name, itype, bc, 1)
+
 template <typename T, int BC, int WM, int WN, int M = 6, int R = 3>
 [[kernel, max_total_threads_per_threadgroup(WM* WN * 32)]] void
 winograd_conv_2d_input_transform(
@@ -445,10 +460,17 @@ winograd_conv_2d_input_transform(
   // Resolve input tile
   constexpr int TH = (A / WM);
   constexpr int TW = (A / WN);
-  int kh = TH * (simd_group_id / WN);
+  const int kh = TH * (simd_group_id / WN);
-  int kw = TW * (simd_group_id % WN);
+  const int kw = TW * (simd_group_id % WN);
-  int bh = M * tid.y + kh;
+  const int bh = M * tid.y + kh - params.pad[1];
-  int bw = M * tid.x + kw;
+  const int bw = M * tid.x + kw - params.pad[0];
+
+  const bool is_edge_w_lo = bw < 0;
+  const bool is_edge_h_lo = bh < 0;
+  const bool is_edge_w_hi = bw + (TW - 1) >= params.iS[0];
+  const bool is_edge_h_hi = bh + (TH - 1) >= params.iS[1];
+  const bool is_edge =
+      is_edge_w_lo || is_edge_h_lo || is_edge_w_hi || is_edge_h_hi;
 
   // Move to the correct input tile
   inp_in += tid.z * params.in_strides[0] + bh * params.in_strides[1] +
@@ -484,10 +506,23 @@ winograd_conv_2d_input_transform(
     for (int h = 0; h < TH; h++) {
       for (int w = 0; w < TW; w++) {
         const device T* in_ptr = inp_in + jump_in[h][w];
+        if (is_edge) {
+          if (((bh + h) < 0 || (bh + h) >= params.iS[1]) ||
+              ((bw + w) < 0 || (bw + w) >= params.iS[0])) {
+            for (int c = simd_lane_id; c < BC; c += 32) {
+              Is[kh + h][kw + w][c] = T(0);
+            }
+          } else {
             for (int c = simd_lane_id; c < BC; c += 32) {
               Is[kh + h][kw + w][c] = in_ptr[c];
             }
           }
+        } else {
+          for (int c = simd_lane_id; c < BC; c += 32) {
+            Is[kh + h][kw + w][c] = in_ptr[c];
+          }
+        }
+      }
     }
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
@@ -652,3 +687,373 @@ winograd_conv_2d_output_transform(
 instantiate_winograd_conv_2d(float32, float);
 instantiate_winograd_conv_2d(bfloat16, bfloat16_t);
 instantiate_winograd_conv_2d(float16, half); // clang-format on
+
+#include "mlx/backend/metal/kernels/steel/attn/mma.h"
+
+template <
+    typename T,
+    bool do_flip = false,
+    int WM = 4,
+    int WN = 1,
+    typename AccumType = float>
+[[kernel]] void winograd_fused(
+    const device T* input [[buffer(0)]],
+    const device T* weight [[buffer(1)]],
+    device T* output [[buffer(2)]],
+    const constant MLXConvParams<2>& params [[buffer(3)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]],
+    uint3 tgp_per_grid [[threadgroups_per_grid]],
+    ushort simd_group_id [[simdgroup_index_in_threadgroup]],
+    ushort simd_lane_id [[thread_index_in_simdgroup]]) {
+  using namespace mlx::steel;
+
+  (void)tgp_per_grid;
+
+  // Winograd F(n x n, r x r)
+  // n x n output window
+  constexpr short FN = 2;
+  // r x r filter size
+  constexpr short FR = 3;
+  // a x a input window, a = n + r - 1
+  constexpr short FA = 4;
+
+  constexpr short kFragSize = 8; // MMA frag size
+
+  constexpr short BT = 8; // Tile block size
+  constexpr short BO = 8; // Output channel block size
+  constexpr short BC = 8; // Input channel block size
+
+  // clang-format off
+  static_assert(BT % (1 * kFragSize) == 0 && 
+                BO % (1 * kFragSize) == 0 && 
+                BC % kFragSize == 0,
+                "Matmuls sizes must be compatible with fragments");
+  // clang-format on
+
+  // Prepare for matmul
+
+  // Warp tile sizes for matmul
+  constexpr short TM = (FA * FA * BT) / (WM * kFragSize);
+  constexpr short TN = (BO) / (WN * kFragSize);
+  constexpr short TK = (BC) / (kFragSize);
+
+  // Warp primitives
+  using MMAFrag_acc_t = BaseMMAFrag<AccumType, kFragSize, kFragSize>;
+
+  // Warp tiles sizes for matmul
+  MMATile<AccumType, 1, TK, MMAFrag_acc_t> Itile;
+  MMATile<AccumType, TK, TN, MMAFrag_acc_t> Wtile;
+  MMATile<AccumType, 1, TN, MMAFrag_acc_t> Otile[TM];
+
+  for (int im = 0; im < 4; im++) {
+    Otile[im].clear();
+  }
+
+  // Threadgroup memory for Weights and Inputs
+  constexpr short BS = BT > BO ? BT : BO;
+  threadgroup T Wt[FA * FA * BC * BO];
+  threadgroup T It[FA * FA * BS * BS];
+
+  // Get thread position in tile
+  short2 simd_coord = MMAFrag_acc_t::get_coord(simd_lane_id);
+  const short sm = simd_coord.y;
+  const short sn = simd_coord.x;
+
+  static_assert(FA * FA * BT == 32 * WM * WN, "Each thread loads one pixel.");
+  const int thr_idx = simd_group_id * 32 + simd_lane_id;
+  const int thr_t = thr_idx / (FA * FA);
+  const int thr_hw = thr_idx % (FA * FA);
+  const int thr_h = thr_hw / FA;
+  const int thr_w = thr_hw % FA;
+
+  // Get batch, tile, and output idx for warp
+  const int b_idx = tid.z;
+  const int t_idx = BT * tid.y + thr_t;
+  const int o_idx = BO * tid.x + thr_t;
+
+  // Divide tile into h, w tile
+  uniform<int> oHu = make_uniform(params.oS[0]);
+  uniform<int> oWu = make_uniform(params.oS[1]);
+  uniform<int> tHu = (oHu + make_uniform(FN - 1)) / make_uniform(FN);
+  uniform<int> tWu = (oWu + make_uniform(FN - 1)) / make_uniform(FN);
+
+  const int oH_idx = FN * (t_idx / tWu);
+  const int oW_idx = FN * (t_idx % tWu);
+  const int iH_idx = oH_idx + thr_h - params.pad[0];
+  const int iW_idx = oW_idx + thr_w - params.pad[1];
+
+  // Move to correct location
+
+  // clang-format off
+  input +=  b_idx * params.in_strides[0] + // N
+           iH_idx * params.in_strides[1] + // H
+           iW_idx * params.in_strides[2];  // W
+
+  weight += o_idx * params.wt_strides[0] + // O
+            thr_h * params.wt_strides[1] + // H 
+            thr_w * params.wt_strides[2];  // W
+  // clang-format on
+
+  // Do edge check prep for input
+  const bool is_edge_w_lo = iH_idx < 0;
+  const bool is_edge_h_lo = iW_idx < 0;
+  const bool is_edge_w_hi = iH_idx >= params.iS[0];
+  const bool is_edge_h_hi = iW_idx >= params.iS[1];
+  const bool is_edge =
+      is_edge_w_lo || is_edge_h_lo || is_edge_w_hi || is_edge_h_hi;
+
+  // Iterate over C
+  for (int c = 0; c < params.C; c += BC) {
+#define tmp_load_wt_idx(o, h, w, c) h* FA* BC* BO + w* BC* BO + c* BO + o
+#define tmp_load_in_idx(t, h, w, c) h* FA* BS* BC + w* BS* BC + t* BC + c
+
+#define tmp_trns_wt_idx(o, h, w, c) h* FA* BC* BO + w* BC* BO + c* BO + o
+#define tmp_trns_in_idx(t, h, w, c) h* FA* BS* BC + w* BS* BC + t* BC + c
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Load weight
+    if (thr_h < FR && thr_w < FR && thr_t < BO) {
+      for (int ic = 0; ic < BC; ic++) {
+        if (do_flip) {
+          Wt[tmp_load_wt_idx(thr_t, FR - 1 - thr_h, FR - 1 - thr_w, ic)] =
+              weight[c + ic];
+        } else {
+          Wt[tmp_load_wt_idx(thr_t, thr_h, thr_w, ic)] = weight[c + ic];
+        }
+      }
+    }
+
+    // Load input
+    if (is_edge) {
+      for (int ic = 0; ic < BC; ic++) {
+        It[tmp_load_in_idx(thr_t, thr_h, thr_w, ic)] = T(0);
+      }
+    } else {
+      for (int ic = 0; ic < BC; ic++) {
+        It[tmp_load_in_idx(thr_t, thr_h, thr_w, ic)] = input[c + ic];
+      }
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Transform weight
+    if (lid.z == 0) {
+      const short ic = lid.y;
+      const short io = lid.x;
+
+      T tmp_0[4][4];
+      T tmp_1[4][4];
+
+      for (int ii = 0; ii < 3; ++ii) {
+        for (int jj = 0; jj < 3; ++jj) {
+          tmp_0[ii][jj] = Wt[tmp_load_wt_idx(io, ii, jj, ic)];
+        }
+      }
+
+      //////////////////////////////////////////////
+
+      tmp_1[0][0] = tmp_0[0][0];
+      tmp_1[0][1] = tmp_0[0][1];
+      tmp_1[0][2] = tmp_0[0][2];
+
+      tmp_1[1][0] = T(0.5) * (tmp_0[0][0] + tmp_0[1][0] + tmp_0[2][0]);
+      tmp_1[1][1] = T(0.5) * (tmp_0[0][1] + tmp_0[1][1] + tmp_0[2][1]);
+      tmp_1[1][2] = T(0.5) * (tmp_0[0][2] + tmp_0[1][2] + tmp_0[2][2]);
+
+      tmp_1[2][0] = tmp_1[1][0] - tmp_0[1][0];
+      tmp_1[2][1] = tmp_1[1][1] - tmp_0[1][1];
+      tmp_1[2][2] = tmp_1[1][2] - tmp_0[1][2];
+
+      tmp_1[3][0] = tmp_0[2][0];
+      tmp_1[3][1] = tmp_0[2][1];
+      tmp_1[3][2] = tmp_0[2][2];
+
+      //////////////////////////////////////////////
+      tmp_0[0][0] = tmp_1[0][0];
+      tmp_0[1][0] = tmp_1[1][0];
+      tmp_0[2][0] = tmp_1[2][0];
+      tmp_0[3][0] = tmp_1[3][0];
+
+      tmp_0[0][1] = T(0.5) * (tmp_1[0][0] + tmp_1[0][1] + tmp_1[0][2]);
+      tmp_0[1][1] = T(0.5) * (tmp_1[1][0] + tmp_1[1][1] + tmp_1[1][2]);
+      tmp_0[2][1] = T(0.5) * (tmp_1[2][0] + tmp_1[2][1] + tmp_1[2][2]);
+      tmp_0[3][1] = T(0.5) * (tmp_1[3][0] + tmp_1[3][1] + tmp_1[3][2]);
+
+      tmp_0[0][2] = tmp_0[0][1] - tmp_1[0][1];
+      tmp_0[1][2] = tmp_0[1][1] - tmp_1[1][1];
+      tmp_0[2][2] = tmp_0[2][1] - tmp_1[2][1];
+      tmp_0[3][2] = tmp_0[3][1] - tmp_1[3][1];
+
+      tmp_0[0][3] = tmp_1[0][2];
+      tmp_0[1][3] = tmp_1[1][2];
+      tmp_0[2][3] = tmp_1[2][2];
+      tmp_0[3][3] = tmp_1[3][2];
+
+      for (int ii = 0; ii < 4; ++ii) {
+        for (int jj = 0; jj < 4; ++jj) {
+          Wt[tmp_trns_wt_idx(io, ii, jj, ic)] = tmp_0[ii][jj];
+        }
+      }
+    }
+
+    // Transform input
+    else {
+      const short it = lid.y;
+      const short ic = lid.x;
+
+      T tmp_0[4][4];
+      T tmp_1[4][4];
+
+      for (int ii = 0; ii < 4; ++ii) {
+        for (int jj = 0; jj < 4; ++jj) {
+          tmp_0[ii][jj] = It[tmp_load_in_idx(it, ii, jj, ic)];
+        }
+      }
+
+      //////////////////////////////////////////////
+
+      tmp_1[0][0] = tmp_0[0][0] - tmp_0[2][0];
+      tmp_1[0][1] = tmp_0[0][1] - tmp_0[2][1];
+      tmp_1[0][2] = tmp_0[0][2] - tmp_0[2][2];
+      tmp_1[0][3] = tmp_0[0][3] - tmp_0[2][3];
+
+      tmp_1[1][0] = tmp_0[1][0] + tmp_0[2][0];
+      tmp_1[1][1] = tmp_0[1][1] + tmp_0[2][1];
+      tmp_1[1][2] = tmp_0[1][2] + tmp_0[2][2];
+      tmp_1[1][3] = tmp_0[1][3] + tmp_0[2][3];
+
+      tmp_1[2][0] = tmp_0[2][0] - tmp_0[1][0];
+      tmp_1[2][1] = tmp_0[2][1] - tmp_0[1][1];
+      tmp_1[2][2] = tmp_0[2][2] - tmp_0[1][2];
+      tmp_1[2][3] = tmp_0[2][3] - tmp_0[1][3];
+
+      tmp_1[3][0] = tmp_0[1][0] - tmp_0[3][0];
+      tmp_1[3][1] = tmp_0[1][1] - tmp_0[3][1];
+      tmp_1[3][2] = tmp_0[1][2] - tmp_0[3][2];
+      tmp_1[3][3] = tmp_0[1][3] - tmp_0[3][3];
+
+      //////////////////////////////////////////////
+      tmp_0[0][0] = tmp_1[0][0] - tmp_1[0][2];
+      tmp_0[1][0] = tmp_1[1][0] - tmp_1[1][2];
+      tmp_0[2][0] = tmp_1[2][0] - tmp_1[2][2];
+      tmp_0[3][0] = tmp_1[3][0] - tmp_1[3][2];
+
+      tmp_0[0][1] = tmp_1[0][1] + tmp_1[0][2];
+      tmp_0[1][1] = tmp_1[1][1] + tmp_1[1][2];
+      tmp_0[2][1] = tmp_1[2][1] + tmp_1[2][2];
+      tmp_0[3][1] = tmp_1[3][1] + tmp_1[3][2];
+
+      tmp_0[0][2] = tmp_1[0][2] - tmp_1[0][1];
+      tmp_0[1][2] = tmp_1[1][2] - tmp_1[1][1];
+      tmp_0[2][2] = tmp_1[2][2] - tmp_1[2][1];
+      tmp_0[3][2] = tmp_1[3][2] - tmp_1[3][1];
+
+      tmp_0[0][3] = tmp_1[0][1] - tmp_1[0][3];
+      tmp_0[1][3] = tmp_1[1][1] - tmp_1[1][3];
+      tmp_0[2][3] = tmp_1[2][1] - tmp_1[2][3];
+      tmp_0[3][3] = tmp_1[3][1] - tmp_1[3][3];
+
+      for (int ii = 0; ii < 4; ++ii) {
+        for (int jj = 0; jj < 4; ++jj) {
+          It[tmp_trns_in_idx(it, ii, jj, ic)] = tmp_0[ii][jj];
+        }
+      }
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Do matmul
+    for (int im = 0; im < 4; im++) {
+      simdgroup_barrier(mem_flags::mem_none);
+      Itile.template load<T, 1, 1, BS, 1>(
+          &It[simd_group_id * FA * BS * BS + im * BS * BS + sm * BS + sn]);
+      simdgroup_barrier(mem_flags::mem_none);
+      Wtile.template load<T, 1, 1, BO, 1>(
+          &Wt[simd_group_id * FA * BC * BO + im * BC * BO + sm * BO + sn]);
+      simdgroup_barrier(mem_flags::mem_none);
+      tile_matmad(Otile[im], Itile, Wtile, Otile[im]);
+    }
+  }
+
+  // Transform and write output
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  for (int im = 0; im < 4; im++) {
+    Otile[im].template store<T, 1, 1, BS, 1>(
+        &It[simd_group_id * FA * BS * BS + im * BS * BS + sm * BS + sn]);
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  if (lid.z == 0) {
+    const short it = lid.y;
+    const short io = lid.x;
+
+    T tmp_0[4][4];
+    T tmp_1[2][4];
+    T tmp_2[2][2];
+
+    for (int ii = 0; ii < 4; ++ii) {
+      for (int jj = 0; jj < 4; ++jj) {
+        tmp_0[ii][jj] = It[tmp_trns_in_idx(it, ii, jj, io)];
+      }
+    }
+
+    tmp_1[0][0] = tmp_0[0][0] + tmp_0[1][0] + tmp_0[2][0];
+    tmp_1[0][1] = tmp_0[0][1] + tmp_0[1][1] + tmp_0[2][1];
+    tmp_1[0][2] = tmp_0[0][2] + tmp_0[1][2] + tmp_0[2][2];
+    tmp_1[0][3] = tmp_0[0][3] + tmp_0[1][3] + tmp_0[2][3];
+
+    tmp_1[1][0] = tmp_0[1][0] - tmp_0[2][0] - tmp_0[3][0];
+    tmp_1[1][1] = tmp_0[1][1] - tmp_0[2][1] - tmp_0[3][1];
+    tmp_1[1][2] = tmp_0[1][2] - tmp_0[2][2] - tmp_0[3][2];
+    tmp_1[1][3] = tmp_0[1][3] - tmp_0[2][3] - tmp_0[3][3];
+
+    tmp_2[0][0] = tmp_1[0][0] + tmp_1[0][1] + tmp_1[0][2];
+    tmp_2[1][0] = tmp_1[1][0] + tmp_1[1][1] + tmp_1[1][2];
+
+    tmp_2[0][1] = tmp_1[0][1] - tmp_1[0][2] - tmp_1[0][3];
+    tmp_2[1][1] = tmp_1[1][1] - tmp_1[1][2] - tmp_1[1][3];
+
+    const int oH_i = FN * ((BT * tid.y + it) / tWu);
+    const int oW_i = FN * ((BT * tid.y + it) % tWu);
+
+    // clang-format off
+    output += b_idx * params.out_strides[0] + // N
+              oH_i * params.out_strides[1] +  // H
+              oW_i * params.out_strides[2] +  // W
+              BO * tid.x;                     // C
+
+    // clang-format on
+
+    output[0 * params.out_strides[1] + 0 * params.out_strides[2] + io] =
+        tmp_2[0][0];
+    output[0 * params.out_strides[1] + 1 * params.out_strides[2] + io] =
+        tmp_2[0][1];
+    output[1 * params.out_strides[1] + 0 * params.out_strides[2] + io] =
+        tmp_2[1][0];
+    output[1 * params.out_strides[1] + 1 * params.out_strides[2] + io] =
+        tmp_2[1][1];
+  }
+}
+
+// clang-format off
+#define instantiate_winograd_conv_2d_fused(name, itype, f)            \
+  template [[host_name("winograd_conv_2d_fused_" #name "_flip" #f)]]  \
+  [[kernel]] void winograd_fused<itype, f>(                           \
+      const device itype* input [[buffer(0)]],                        \
+      const device itype* weight [[buffer(1)]],                       \
+      device itype* output [[buffer(2)]],                             \
+      const constant MLXConvParams<2>& params [[buffer(3)]],          \
+      uint3 tid [[threadgroup_position_in_grid]],                     \
+      uint3 lid [[thread_position_in_threadgroup]],                   \
+      uint3 tgp_per_grid [[threadgroups_per_grid]],                   \
+      ushort simd_group_id [[simdgroup_index_in_threadgroup]],        \
+      ushort simd_lane_id [[thread_index_in_simdgroup]]);
+
+#define instantiate_winograd_conv_2d_fused_2(name, itype) \
+  instantiate_winograd_conv_2d_fused(name, itype, 0)      \
+  instantiate_winograd_conv_2d_fused(name, itype, 1)
+
+instantiate_winograd_conv_2d_fused_2(float32, float);
+instantiate_winograd_conv_2d_fused_2(float16, float16_t);
+instantiate_winograd_conv_2d_fused_2(bfloat16, bfloat16_t);
+// clang-format on

mlx/ops.cpp

@@ -3882,7 +3882,7 @@ array conv_general(
 
   return array(
       std::move(out_shape),
-      in.dtype(),
+      out_type,
       std::make_shared<Convolution>(
           to_stream(s),
           stride,

python/tests/test_conv.py

@@ -341,7 +341,7 @@ class TestConv(mlx_tests.MLXTestCase):
                     atol, rtol = 1e-1, 1e-3
                 else:
                     atol, rtol = 1e-5, 1e-6
-                self.assertTrue(np.allclose(out_pt, out_mx, atol=atol))
+                self.assertTrue(np.allclose(out_pt, out_mx, atol=atol, rtol=rtol))
 
         for dtype in ("float32", "bfloat16"):
             for N, C, O in (