merge from public repo

include/ck_tile/ops/smoothquant/kernel/smoothquant_kernel.hpp

@@ -19,7 +19,8 @@ struct SmoothquantHostArgs
 
     index_t m;
     index_t n;
-    index_t stride; // row_stride
+    index_t x_stride; // input row_stride
+    index_t y_stride; // output row_stride
 };
 
 // TODO: Extract some type to wrapper class
@@ -58,14 +59,21 @@ struct Smoothquant
 
         index_t m;
         index_t n;
-        index_t stride; // row_stride
+        index_t x_stride; // input row_stride
+        index_t y_stride; // out row_stride
     };
     using Hargs = SmoothquantHostArgs;
 
     CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
     {
-        return Kargs{
-            hargs.p_x, hargs.p_xscale, hargs.p_yscale, hargs.p_qy, hargs.m, hargs.n, hargs.stride};
+        return Kargs{hargs.p_x,
+                     hargs.p_xscale,
+                     hargs.p_yscale,
+                     hargs.p_qy,
+                     hargs.m,
+                     hargs.n,
+                     hargs.x_stride,
+                     hargs.y_stride};
     }
 
     CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
@@ -116,7 +124,7 @@ struct Smoothquant
             const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                 static_cast<const XDataType*>(kargs.p_x),
                 make_tuple(kargs.m, kargs.n),
-                make_tuple(kargs.stride, 1),
+                make_tuple(kargs.x_stride, 1),
                 number<Vector_N>{},
                 number<1>{});
 
@@ -157,7 +165,7 @@ struct Smoothquant
             auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                 static_cast<QYDataType*>(kargs.p_qy),
                 make_tuple(kargs.m, kargs.n),
-                make_tuple(kargs.stride, 1),
+                make_tuple(kargs.y_stride, 1),
                 number<Vector_N>{},
                 number<1>{});
 

include/ck_tile/ref/README.md

+# reference
+
+this folder contains reference implementation of a specific op. Note by including a specific header, you are including the implementation(expecially the gpu implementation) into your source code, and compile that kernel into the fatbin, hence may increase your kernel obj code length. Usually the header starts with `reference_` is a cpu reference implementation. The header starts with `naive_` contains a gpu implementation with a small launcher.
+
+TODO: move `host/reference` under this folder

include/ck_tile/remod.py

@@ -7,6 +7,7 @@ import copy
 
 NS = 'ck_tile'
 OPS = 'ops'
+REF = 'ref'
 OPS_COMMON = 'common' # common header will be duplicated into ops/* other module
 
 HEADER_COMMON = f"""// SPDX-License-Identifier: MIT
@@ -29,6 +30,9 @@ class submodule_t:
     def push(self, f):
         if len(f.parents) != 1: # ignore ./xxx.hpp
             mod = get_module(f)
+            # ref is supposed to include one header on demand
+            if mod == REF:
+                return
             if mod == OPS:
                 if mod not in self.m.keys():
                     self.m[mod] = dict()

library/src/tensor_operation_instance/gpu/gemm_universal_batched/device_batched_gemm_xdl_universal_bf16_bf16_bf16/device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp

@@ -52,6 +52,9 @@ using device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_instances =
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   224,   256,    64,   8,   8,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           2,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   224,    64,   8,   8,  16,   16,    8,    7,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          2,           1,                   S<1, 32, 1, 8>,          S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   160,    64,   8,   8,  16,   16,    8,    5,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          2,           1,                   S<1, 32, 1, 8>,          S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   160,    64,   8,   8,  32,   32,    1,    5,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 64, 1, 4>,          S<8>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+        DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   160,   128,    64,   8,   8,  32,   32,    5,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,     Row,     BF16,   BF16, DsDataType,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,         S<4>,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>

library/src/tensor_operation_instance/gpu/gemm_universal_batched/device_batched_gemm_xdl_universal_f8_f8_bf16/device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp

@@ -42,6 +42,7 @@ using device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std
         //##################################|        |        |         |        | Type|  Type|         Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
         //##################################|        |        |         |        |     |      |             |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
         //##################################|        |        |         |        |     |      |             |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+        
 #ifdef __gfx94__
         // Compute friendly
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,    Row,     F8,     F8, DsDataType, BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    64,  16,  16,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,               S<8>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
@@ -72,6 +73,7 @@ using device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_instances = std:
         //##################################|        |        |         |        | Type|  Type|         Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|          Pipeline|                     Pipeline|
         //##################################|        |        |         |        |     |      |             |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|         Scheduler|                     Verision|
         //##################################|        |        |         |        |     |      |             |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
+
 #if defined(__gfx94__) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,    Row,     F8,     F8, DsDataType,   BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    128, 16,  16,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               S<2>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceBatchedGemmMultiD_Xdl_CShuffle_V3<  Row,     Col, DsLayout,    Row,     F8,     F8, DsDataType,   BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    128, 16,  16,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,               S<4>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,

profiler/include/profiler/profile_gemm_universal_impl.hpp

@@ -144,6 +144,7 @@ bool profile_gemm_universal_impl(int do_verification,
     }
 
     std::string best_op_name;
+    std::optional<std::string> best_op_object_name;
     float best_ave_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
@@ -225,7 +226,8 @@ bool profile_gemm_universal_impl(int do_verification,
                     }
                 }
 
-                std::string op_name = op_ptr->GetTypeString();
+                std::string op_name                    = op_ptr->GetTypeString();
+                std::optional<std::string> op_obj_name = op_ptr->GetObjectName();
 
                 float ave_time = invoker_ptr->Run(argument_ptr.get(),
                                                   StreamConfig{nullptr,
@@ -251,11 +253,12 @@ bool profile_gemm_universal_impl(int do_verification,
 
                 if(tflops > best_tflops && ave_time > 1e-10)
                 {
-                    best_op_name    = op_name;
-                    best_tflops     = tflops;
-                    best_ave_time   = ave_time;
-                    best_gb_per_sec = gb_per_sec;
-                    best_kbatch     = kbatch_curr;
+                    best_op_name        = op_name;
+                    best_op_object_name = op_obj_name;
+                    best_tflops         = tflops;
+                    best_ave_time       = ave_time;
+                    best_gb_per_sec     = gb_per_sec;
+                    best_kbatch         = kbatch_curr;
                 }
             }
             else
@@ -306,6 +309,9 @@ bool profile_gemm_universal_impl(int do_verification,
               << " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
               << " GB/s, " << best_op_name << std::endl;
 
+    if(best_op_object_name)
+        std::cout << best_op_object_name.value() << std::endl;
+
     return pass;
 }