optimize for dgrad

example/ck_tile/02_layernorm2d/instances/layernorm2d_bwd_bf16_n64_n128_instance.cpp

@@ -28,6 +28,10 @@
 // large n
 // template float layernorm2d_bwd_<trait_<ck_tile::bf16_t, 1,  32,  4,  16, 8,  true>>(const S&, A);
 // template float layernorm2d_bwd_<trait_<ck_tile::fp16_t, 1,  32,  4,  16, 8,  true>>(const S&, A);
-template float layernorm2d_bwd_<trait_<ck_tile::bf16_t, 1,  4,  1,  128, 8,  true>>(const S&, A);
-template float layernorm2d_bwd_<trait_<ck_tile::fp16_t, 1,  4,  1,  128, 8,  true>>(const S&, A);
+template float layernorm2d_bwd_<trait_<ck_tile::bf16_t, 1,  4,  1,  64, 8,  true>>(const S&, A);
+template float layernorm2d_bwd_<trait_<ck_tile::fp16_t, 1,  4,  1,  64, 8,  true>>(const S&, A);
+
+// two pass
+template float layernorm2d_bwd_<trait_<ck_tile::bf16_t, 1,  4,  2,  32, 8,  true>>(const S&, A);
+template float layernorm2d_bwd_<trait_<ck_tile::fp16_t, 1,  4,  2,  32, 8,  true>>(const S&, A);
 // clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_bwd_instance_common.hpp

@@ -27,7 +27,7 @@ float layernorm2d_bwd_(const S& s, A a)
         typename Traits_::Shape,
         Traits_::kPadN>;
 
-    using Pipeline = ck_tile::Layernorm2dBwdGammaBetaPipeline<PipelineProblem>;
+    using Pipeline = ck_tile::Layernorm2dBwdGammaBetaPipelineTwoPass<PipelineProblem>;
 
     using Kernel = ck_tile::Layernorm2dBwdGammaBeta<Pipeline>;
 

example/ck_tile/02_layernorm2d/layernorm2d_bwd.cpp

@@ -139,7 +139,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
         traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
 
     std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(GammaDataType) * n +
-                           sizeof(BetaDataType) * n + sizeof(YDataType) * m * n;
+                           sizeof(MeanDataType) * m + sizeof(InvStdDataType) * m + sizeof(YDataType) * m * n + sizeof(XDataType);
 
     float gb_per_sec = num_byte / 1.E6 / ave_time;
     std::cout << sizeof(ComputeDataType) << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;

example/ck_tile/02_layernorm2d/layernorm2d_bwd.hpp

@@ -136,7 +136,7 @@ struct layernorm2d_bwd_b16_
     /* data */
     //using Trait = trait_<data_type,   1,  1,  1,  256,  1,  true>;
     //using Trait = trait_<data_type,   1,  8,  64,  4,  8,  true>;
-    using Trait = trait_<data_type,   1,  4,  1,  128, 8,  true>;
+    using Trait = trait_<data_type,   1,  4,  1,  64, 8,  true>;
     float operator() (layernorm2d_bwd_traits /*t*/,
                       layernorm2d_bwd_args a,
                       const ck_tile::stream_config& s) {

include/ck_tile/ops/layernorm2d.hpp

@@ -14,6 +14,7 @@
 #include "ck_tile/ops/layernorm2d/kernel/layernorm2d_bwd_gamma_beta_kernel.hpp"
 #include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_default_policy.hpp"
 #include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_gamma_beta.hpp"
+#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_two_pass_dx.hpp"
 #include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_problem.hpp"
 
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/layernorm2d/pipeline/2passtmp → include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_two_pass_dx.hpp

@@ -12,7 +12,7 @@
 namespace ck_tile {
 
 template <typename Problem_, typename Policy_ = Layernorm2dBwdGammaBetaPipelineDefaultPolicy>
-struct Layernorm2dBwdGammaBetaPipeline
+struct Layernorm2dBwdGammaBetaPipelineTwoPass
 {
     using Problem      = ck_tile::remove_cvref_t<Problem_>;
     using Policy       = ck_tile::remove_cvref_t<Policy_>;
@@ -33,7 +33,8 @@ struct Layernorm2dBwdGammaBetaPipeline
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
     {
-        return Policy::template GetSmemSize<Problem>();
+        return ReducePolicy::template GetSmemSize<Problem>();
+        //GetBlockReduce2dCrossWarpSync<Problem>().template GetSmemSize<y_block_tile>();
     }
     template <typename XWindow,
               typename GammaWindow,
@@ -62,18 +63,17 @@ struct Layernorm2dBwdGammaBetaPipeline
                                    ck_tile::index_t row_size,
                                    void* smem) const
     {
-        (void)smem;
-
+        
         auto gamma_beta_dist  = Policy::template MakeGammaBetaBlockTileDistribution<Problem>();
         auto dgamma_beta_dist = Policy::template MakeDGammaBetaBlockTileDistribution<Problem>();
         auto mean_dist        = Policy::template MakeMeanBlockTileDistribution<Problem>();
         auto x_dist           = Policy::template MakeXBlockTileDistribution<Problem>();
 
-        const auto x_window       = make_tile_window(x_window_, x_dist);
-        const auto dy_window      = make_tile_window(dy_window_, x_dist);
-        const auto gamma_window   = make_tile_window(gamma_window_, gamma_beta_dist); // TO CHECK
-        const auto mean_window    = make_tile_window(mean_window_, mean_dist);
-        const auto inv_std_window = make_tile_window(inv_std_window_, mean_dist);
+        auto x_window       = make_tile_window(x_window_, x_dist);
+        auto dy_window      = make_tile_window(dy_window_, x_dist);
+        auto gamma_window   = make_tile_window(gamma_window_, gamma_beta_dist); // TO CHECK
+        auto mean_window    = make_tile_window(mean_window_, mean_dist);
+        auto inv_std_window = make_tile_window(inv_std_window_, mean_dist);
 
         auto dgamma_window = make_tile_window(dgamma_window_, dgamma_beta_dist);
         auto dbeta_window  = make_tile_window(dbeta_window_, dgamma_beta_dist);
@@ -92,11 +92,11 @@ struct Layernorm2dBwdGammaBetaPipeline
         clear_tile(ds_tile);
         clear_tile(db_tile);
 
-        auto dgamma_tile = make_static_distributed_tensor<GammaDataType>(dgamma_beta_dist);
-        auto dbeta_tile  = make_static_distributed_tensor<BetaDataType>(dgamma_beta_dist);
+        // auto dgamma_tile = make_static_distributed_tensor<GammaDataType>(dgamma_beta_dist);
+        // auto dbeta_tile  = make_static_distributed_tensor<BetaDataType>(dgamma_beta_dist);
         auto dx_tile     = make_static_distributed_tensor<XDataType>(x_dist);
-        auto dgamma      = cast_tile<ComputeDataType>(dgamma_tile);
-        auto dbeta       = cast_tile<ComputeDataType>(dbeta_tile);
+        // auto dgamma      = cast_tile<ComputeDataType>(dgamma_tile);
+        // auto dbeta       = cast_tile<ComputeDataType>(dbeta_tile);
         auto dx          = cast_tile<ComputeDataType>(dx_tile);
 
         static constexpr index_t Block_N = Problem::BlockShape::Block_N;
@@ -119,36 +119,48 @@ struct Layernorm2dBwdGammaBetaPipeline
                 const auto gamma     = type_convert<ComputeDataType>(gamma_tile[j_idx]);
                 ds_tile(i_idx) += dy * gamma * x;
                 db_tile(i_idx) += dy * gamma;
-                // printf("threadidx=%d, blockidx=%d, ds_tile=%f\n",threadIdx.x, blockIdx.x, ds_tile[i_idx]);
+                // dx(idx) = dy * gamma;
+                // ds_tile(i_idx) += dx[idx] * x;
+                // db_tile(i_idx) += dx[idx];
+            // printf("db_tile pre: threadidx=%d, blockidx=%d, db_tile=%f\n",threadIdx.x, blockIdx.x, db_tile[i_idx]);
+            // printf("dy_tile: threadidx=%d, blockidx=%d, dy_tile=%f\n",threadIdx.x, blockIdx.x, dy);
+            // printf("x: threadidx=%d, blockidx=%d, x_tile=%f\n",threadIdx.x, blockIdx.x, x);
+            // printf("gamma: threadidx=%d, blockidx=%d, gamma_tile=%f\n",threadIdx.x, blockIdx.x, gamma);
             });
         }
 
         auto block_reduce2d_sync = ReducePolicy::template GetBlockReduce2dSync<Problem>();
+        auto block_reduce2d_cross_warp_sync = ReducePolicy::template GetBlockReduce2dCrossWarpSync<Problem>();
         block_reduce2d_sync(ds_tile, ck_tile::ReduceOp::Add{});
         block_reduce2d_sync(db_tile, ck_tile::ReduceOp::Add{});
+        // block_reduce2d_cross_warp_sync(ds_tile, smem, ck_tile::ReduceOp::Add{});
+        // block_reduce2d_cross_warp_sync(db_tile, smem, ck_tile::ReduceOp::Add{});
 
         // sweep_tile(x_tile, [&](auto idx) {
         //     constexpr auto i_idx = make_tuple(idx[number<0>{}]);
-        //     printf("post::threadidx=%d, blockidx=%d, ds_tile=%f\n",threadIdx.x, blockIdx.x,
-        //     ds_tile[i_idx]);
+        //     printf("db_tile post: threadidx=%d, blockidx=%d, db_tile=%f\n",threadIdx.x, blockIdx.x,
+        //     db_tile[i_idx]);
         // });
 
-        //store_tile(ds_window, ds_tile);
-        //store_tile(db_window, db_tile);
+        // store_tile(ds_window, ds_tile);
+        // store_tile(db_window, db_tile);
 
         ck_tile::index_t stride_to_right_most_window = row_size % Block_N == 0 ? row_size - Block_N : row_size - row_size % Block_N;
         move_tile_window(x_window, {0, -Block_N});
         move_tile_window(dy_window, {0, -Block_N});
         move_tile_window(gamma_window, {-Block_N});
         move_tile_window(dx_window, {0, stride_to_right_most_window});
-        move_tile_window(dbeta_window, {0, stride_to_right_most_window});
-        move_tile_window(dgamma_window, {0, stride_to_right_most_window});
+        // move_tile_window(dbeta_window, {0, stride_to_right_most_window});
+        // move_tile_window(dgamma_window, {0, stride_to_right_most_window});
 
         using XDistributedTensor = decltype(load_tile(x_window));
         constexpr auto spans     = XDistributedTensor::get_distributed_spans();
 
         for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
         {
+            const auto x_tile         = load_tile(x_window);
+            const auto dy_tile        = load_tile(dy_window);
+            const auto gamma_tile     = load_tile(gamma_window);
             sweep_tile_span(spans[number<0>{}], [&](auto i_idx) {
                 constexpr auto idx0 = make_tuple(i_idx);
                 const auto mean     = type_convert<ComputeDataType>(mean_tile[idx0]);
@@ -157,26 +169,28 @@ struct Layernorm2dBwdGammaBetaPipeline
                 auto c = -b * mean - db_tile[idx0] * inv_std / row_size;
 
                 sweep_tile_span(spans[number<1>{}], [&](auto j_idx) {
+                constexpr auto idx1   = make_tuple(j_idx);
                     constexpr auto idx    = make_tuple(i_idx, j_idx);
-                    constexpr auto gb_idx = make_tuple(number<0>{}, j_idx);
+                    //constexpr auto gb_idx = make_tuple(number<0>{}, j_idx);
                     const auto x          = type_convert<ComputeDataType>(x_tile[idx]);
                     const auto dy         = type_convert<ComputeDataType>(dy_tile[idx]);
-                    const auto gamma      = type_convert<ComputeDataType>(gamma_tile[idx]);
-                    dbeta(gb_idx) += dy;
-                    dgamma(gb_idx) += dy * (x - mean) * inv_std;
+                    const auto gamma      = type_convert<ComputeDataType>(gamma_tile[idx1]);
+                    // dbeta(gb_idx) += dy;
+                    // dgamma(gb_idx) += dy * (x - mean) * inv_std;
                     dx(idx) = dy * gamma * inv_std + b * x + c;
+                    //printf("dx: threadidx=%d, blockidx=%d, dx_tile=%f\n",threadIdx.x, blockIdx.x, dx(idx));
                 });
             });
-            store_tile(dbeta_window, cast_tile<BetaDataType>(dbeta));
-            store_tile(dgamma_window, cast_tile<GammaDataType>(dgamma));
+            // store_tile(dbeta_window, cast_tile<BetaDataType>(dbeta));
+            // store_tile(dgamma_window, cast_tile<GammaDataType>(dgamma));
             store_tile(dx_window, cast_tile<XDataType>(dx));
 
             move_tile_window(x_window, {0, -Block_N});
             move_tile_window(dy_window, {0, -Block_N});
             move_tile_window(gamma_window, {-Block_N});
             move_tile_window(dx_window, {0, -Block_N});
-            move_tile_window(dbeta_window, {0, -Block_N});
-            move_tile_window(dgamma_window, {0, -Block_N});
+            // move_tile_window(dbeta_window, {0, -Block_N});
+            // move_tile_window(dgamma_window, {0, -Block_N});
         }
     }
 };