Adding seed and offset pointer support to the philox random number generator. (#1523)

* Adding seed and offset pointer support to the philox random number generator.

* Separating seed and offset pointer checks with different condition statements.

* Changes include, adding support for device seed and offset pointers, union is used to store seed/offset values and device pointers to minimize device SGPRs.

* Correcting a typo in the readme file

* Re-format files using remod.py

* Use STL type for API parameters

* Use simpler struct design for drop_seed & drop_offset

* Undo unnecessary changes

* Sync kargs style for fmha_fwd.hpp/.cpp

* Use templated union to reduce code

* Use structured binding to make code more readable

---------

Co-authored-by: Sudhir Kylasa <sukylasa@amd.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>

[ROCm/composable_kernel commit: c24fae2346]

example/ck_tile/01_fmha/README.md

@@ -70,8 +70,13 @@ args:
        -seed    random seed used for initializing input tensors. 0 for non-deterministic seed (default:11939)
      -warmup    number of iterations before benchmark the kernel (default:5)
      -repeat    number of iterations to benchmark the kernel (default:20)
+  -drop_seed    seed for the random number generator for the dropout layer, default is 1
+-drop_offset    offset for the dropout layer which is used during random number generation, default is 0
+ -drop_prefs    flag to indicate `drop_seed` and `drop_offset` values if present on the GPU, default is 0, 0 - host, 1 - GPU
 ```
-Example: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
+Example 1: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
+Example 2: `./bin/tile_example_fmha_fwd -b=1 -h=8 -s=16384 -d=64 -drop_prefs=1 -drop_seed=10 -drop_offset=1234` will run a fmha case with 
+  batch=1, nhead=8, sequence length=16384, hdim=64, drop_seed=0 (in GPU memory), drop_offset=1234 (in GPU memory) fp16 case
 
 ## support features
 Currently we are still in rapid development stage, so more features/optimizations will be coming soon.

example/ck_tile/01_fmha/fmha_bwd.cpp

@@ -85,6 +85,9 @@ auto create_args(int argc, char* argv[])
         .insert("p_drop", "0", "0~1 probability of dropout")
         .insert("drop_seed", "1", "seed for random number generator")
         .insert("drop_offset", "0", "offset for random number generator")
+        .insert("drop_prefs",
+                "0",
+                "seed and offset values are present on GPU; 0 - host, 1 - device/GPU")
         .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
         .insert("warmup", "5", "number of iterations before benchmark the kernel")
         .insert("repeat", "20", "number of iterations to benchmark the kernel")
@@ -158,6 +161,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     float p_drop         = arg_parser.get_float("p_drop");
     uint64_t drop_seed   = arg_parser.get_uint64("drop_seed");
     uint64_t drop_offset = arg_parser.get_uint64("drop_offset");
+    bool drop_prefs      = arg_parser.get_bool("drop_prefs");
+
     if(use_dbias && bias.type != bias_enum::elementwise_bias)
     {
         std::cerr << "dbias only exists when bias type is elementwise" << std::endl;
@@ -381,6 +386,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     ck_tile::DeviceMem dbias_buf(dbias_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
     ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
+    ck_tile::DeviceMem drop_seed_buf(drop_prefs ? sizeof(uint64_t) : 0);
+    ck_tile::DeviceMem drop_offset_buf(drop_prefs ? sizeof(uint64_t) : 0);
     ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem dq_acc_buf(dq_acc_host.get_element_space_size_in_bytes());
 
@@ -391,6 +398,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     do_buf.ToDevice(do_host.data());
     seqstart_q.ToDevice(seqstart_q_host.data());
     seqstart_k.ToDevice(seqstart_k_host.data());
+    drop_seed_buf.ToDevice(drop_prefs ? &drop_seed : nullptr);
+    drop_offset_buf.ToDevice(drop_prefs ? &drop_offset : nullptr);
     alibi_slope_buf.ToDevice(alibi_slope_host.data());
 
     // clang-format off
@@ -472,6 +481,18 @@ bool run(const ck_tile::ArgParser& arg_parser)
         const ck_tile::index_t split_stride_dq_acc =
             (shape_batch * nhead * shape_seqlen_q * hdim_q);
 
+        const auto drop_seed_offset = [&]() -> decltype(fmha_bwd_args::drop_seed_offset) {
+            if(drop_prefs)
+            {
+                return std::make_pair(drop_seed_buf.GetDeviceBuffer(),
+                                      drop_offset_buf.GetDeviceBuffer());
+            }
+            else
+            {
+                return std::make_pair(drop_seed, drop_offset);
+            }
+        }();
+
         return fmha_bwd_args{q_buf.GetDeviceBuffer(),
                              k_buf.GetDeviceBuffer(),
                              v_buf.GetDeviceBuffer(),
@@ -545,7 +566,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              static_cast<ck_tile::index_t>(mask.type),
                              p_drop,
                              p_undrop,
-                             {drop_seed, drop_offset}};
+                             drop_seed_offset};
     }();
 
     float ave_time = fmha_bwd(fmha_traits, fmha_args, stream_config);

example/ck_tile/01_fmha/fmha_bwd.hpp

@@ -9,7 +9,10 @@
 #include "ck_tile/ops/epilogue.hpp"
 #include "mask.hpp"
 #include "bias.hpp"
+
 #include <type_traits>
+#include <utility>
+#include <variant>
 
 template <typename DataType>
 struct FmhaBwdTypeConfig;
@@ -135,7 +138,8 @@ struct fmha_bwd_args
     ck_tile::index_t mask_type;
     float p_drop;
     float p_undrop;
-    std::tuple<uint64_t, uint64_t> drop_seed_offset;
+    std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+        drop_seed_offset;
 };
 
 template <typename FmhaBwdDQDKDVKernel>

example/ck_tile/01_fmha/fmha_fwd.cpp

@@ -122,6 +122,9 @@ auto create_args(int argc, char* argv[])
         .insert("p_drop", "0", "0~1 probability of dropout")
         .insert("drop_seed", "1", "seed for random number generator")
         .insert("drop_offset", "0", "offset for random number generator")
+        .insert("drop_prefs",
+                "0",
+                "seed and offset values are present on GPU; 0 - host, 1 - device/GPU")
         .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
         .insert(
             "rotary_dim", "0", "RoPE rotary dimension. rotary_dim <= 0 means not apply RoPE at all")
@@ -442,6 +445,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     float p_drop         = arg_parser.get_float("p_drop");
     uint64_t drop_seed   = arg_parser.get_uint64("drop_seed");
     uint64_t drop_offset = arg_parser.get_uint64("drop_offset");
+    bool drop_prefs      = arg_parser.get_bool("drop_prefs");
+
     if(p_drop < 0.0f || p_drop > 1.0f)
     {
         std::cerr << "The value of p_drop should be 0~1" << std::endl;
@@ -756,6 +761,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
         need_append_kvcache ? cache_seqlen_ks.size() * sizeof(int32_t) : 0);
     ck_tile::DeviceMem rotary_cos_buf(rotary_cos_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem rotary_sin_buf(rotary_sin_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem drop_seed_buf(drop_prefs ? sizeof(uint64_t) : 0);
+    ck_tile::DeviceMem drop_offset_buf(drop_prefs ? sizeof(uint64_t) : 0);
     ck_tile::DeviceMem randval_buf(randval_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem block_table_buf(block_table_host.get_element_space_size_in_bytes());
@@ -774,6 +781,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     cache_seqlen_k_buf.ToDevice(need_append_kvcache ? cache_seqlen_ks.data() : nullptr);
     rotary_cos_buf.ToDevice(rotary_cos_host.data());
     rotary_sin_buf.ToDevice(rotary_sin_host.data());
+    drop_seed_buf.ToDevice(drop_prefs ? &drop_seed : nullptr);
+    drop_offset_buf.ToDevice(drop_prefs ? &drop_offset : nullptr);
     alibi_slope_buf.ToDevice(alibi_slope_host.data());
     block_table_buf.ToDevice(block_table_host.data());
     cache_batch_idx_buf.ToDevice(cache_batch_idx_host.data());
@@ -1013,9 +1022,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
                 args.nhead_stride_randval = nhead_stride_randval;
                 args.batch_stride_randval = batch_stride_randval;
 
-                args.p_drop           = p_drop;
-                args.s_randval        = s_randval;
-                args.drop_seed_offset = std::tie(drop_seed, drop_offset);
+                args.p_drop    = p_drop;
+                args.s_randval = s_randval;
+                if(drop_prefs)
+                {
+                    args.drop_seed_offset = std::make_pair(drop_seed_buf.GetDeviceBuffer(),
+                                                           drop_offset_buf.GetDeviceBuffer());
+                }
+                else
+                {
+                    args.drop_seed_offset = std::make_pair(drop_seed, drop_offset);
+                }
             }
             else if constexpr(std::is_same_v<fmha_fwd_splitkv_args, std::decay_t<decltype(args)>>)
             {

example/ck_tile/01_fmha/fmha_fwd.hpp

@@ -13,6 +13,8 @@
 #include "rotary.hpp"
 
 #include <type_traits>
+#include <utility>
+#include <variant>
 
 template <typename DataType>
 struct FmhaFwdTypeConfig;
@@ -144,7 +146,9 @@ struct fmha_fwd_args
 
     float p_drop;
     bool s_randval;
-    std::tuple<uint64_t, uint64_t> drop_seed_offset;
+
+    std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+        drop_seed_offset;
 };
 
 struct fmha_fwd_splitkv_args