Merge branch 'develop' into add_fp16_wmma_conv_instance

include/ck/utility/reduction_common.hpp

@@ -25,16 +25,4 @@ struct float_equal_zero
     };
 };
 
-template <index_t N>
-static constexpr __device__ index_t get_shift()
-{
-    return (get_shift<N / 2>() + 1);
-};
-
-template <>
-constexpr __device__ index_t get_shift<1>()
-{
-    return (0);
-}
-
 } // namespace ck

include/ck/utility/type_convert.hpp

@@ -62,24 +62,6 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, half_t>(half_
     return type_convert<bhalf_t>(x_fp32);
 }
 
-// convert bfp16 to int32 via fp32
-template <>
-inline __host__ __device__ constexpr int32_t type_convert<int32_t, bhalf_t>(bhalf_t x)
-{
-    float x_fp32 = type_convert<float>(x);
-
-    return static_cast<int32_t>(x_fp32);
-}
-
-// convert int32 to bfp16 via fp32
-template <>
-inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int32_t>(int32_t x)
-{
-    float x_fp32 = static_cast<float>(x);
-
-    return type_convert<bhalf_t>(x_fp32);
-}
-
 // convert bfp16 to int8 via fp32
 template <>
 inline __host__ __device__ constexpr int8_t type_convert<int8_t, bhalf_t>(bhalf_t x)

include/ck/utility/workgroup_synchronization.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "ck/host_utility/hip_check_error.hpp"
+
+namespace ck {
+
+// Initialization flag of Barrier object, can be any value except for zero
+static constexpr int BarrierInitFlag = 0x7856;
+
+// 1) only the first thread-block in the synchronizaton group is supposed to call this function. It
+// is the responsibility of the user to ensure the two integer values in p_control_bits are zeros
+// before calling gms_init().
+// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
+// repetitious initialization of p_control_bits buffer is required
+static __device__ void gms_init(int NumWarps, int* p_control_bits)
+{
+    union
+    {
+        int two32[2];
+        unsigned long one64;
+    } regs;
+
+    regs.two32[0] = BarrierInitFlag;
+    regs.two32[1] = NumWarps;
+
+    if(threadIdx.x == 0)
+        atomicCAS(reinterpret_cast<unsigned long*>(p_control_bits), 0, regs.one64);
+};
+
+// all the workgroups in the synchronization group is supposed to call this function
+static __device__ void gms_barrier(int* p_control_bits)
+{
+    constexpr int mask = warpSize - 1;
+
+    if((threadIdx.x & mask) == 0)
+    {
+        // ensure the barrier object is initialized
+        do
+        {
+            const int r0 = __atomic_load_n(&p_control_bits[0], __ATOMIC_RELAXED);
+
+            if(r0 == BarrierInitFlag)
+                break;
+
+        } while(true);
+
+        // go ahead toward the barrier line
+        atomicSub(&p_control_bits[1], 1);
+
+        // wait until all warps have arrived
+        do
+        {
+            const int r1 = __atomic_load_n(&p_control_bits[1], __ATOMIC_RELAXED);
+
+            if(r1 == 0)
+                break;
+
+        } while(true);
+    };
+};
+
+// 1) Only the first thread-block in the synchronizaton group is supposed to call this function.
+// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
+// repetitious initialization of p_control_bits buffer is required
+static __device__ void gms_reset(int* p_control_bits)
+{
+    // reset the barrier object
+    if(threadIdx.x == 0)
+        (void)atomicCAS(&p_control_bits[0], BarrierInitFlag, 0);
+};
+
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp

@@ -39,7 +39,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
     std::vector<std::unique_ptr<
         DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#ifdef __int8__
 void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<1,
                                                   NWC,
@@ -51,7 +51,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#endif
 // conv2d backward data
 void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
@@ -88,7 +88,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#ifdef __int8__
 void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
                                                   NHWC,
@@ -100,7 +100,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#endif
 // conv2d dl
 void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
@@ -125,7 +125,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#ifdef __int8__
 void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
                                                   NHWC,
@@ -137,6 +137,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
+#endif
 // conv3d backward data
 void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<3,
@@ -173,7 +174,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#ifdef __int8__
 void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<3,
                                                   NDHWC,
@@ -185,7 +186,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#endif
 template <ck::index_t NumDimSpatial,
           typename InLayout,
           typename WeiLayout,
@@ -239,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             {
                 add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(op_ptrs);
             }
+#ifdef __int8__
             else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
                               is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(op_ptrs);
             }
+#endif
         }
         else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
                           is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
@@ -266,12 +269,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
             }
+#ifdef __int8__
             else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
                               is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
                 add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
             }
+#endif
         }
         else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
                           is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
@@ -292,11 +297,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             {
                 add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(op_ptrs);
             }
+#ifdef __int8__
             else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
                               is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(op_ptrs);
             }
+#endif
         }
 
         return op_ptrs;

library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp

@@ -77,7 +77,7 @@ void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#ifdef __int8__
 void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
@@ -118,6 +118,27 @@ void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(
         DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+
 void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
@@ -183,26 +204,6 @@ void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
         DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
-    std::vector<std::unique_ptr<
-        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
-        instances);
-
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
-    std::vector<std::unique_ptr<
-        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
-        instances);
-
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
-    std::vector<std::unique_ptr<
-        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
-        instances);
-
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
-    std::vector<std::unique_ptr<
-        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
-        instances);
-
 void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
@@ -388,6 +389,7 @@ struct DeviceOperationInstanceFactory<
                 add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
             }
         }
+#ifdef __int8__
         else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
                           is_same_v<CDataType, int8_t>)
         {
@@ -420,7 +422,7 @@ struct DeviceOperationInstanceFactory<
                 add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(op_ptrs);
             }
         }
-
+#endif
         return op_ptrs;
     }
 };

library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <cstdlib>
 

library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <cstdlib>
 

library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <cstdlib>