Merge branch 'develop' into ck_host_lib

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instance.cpp

@@ -25,7 +25,7 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances<
             3,
             NDHWGC,
             GKZYXC,

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_instance.cpp

@@ -25,7 +25,7 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances<
             3,
             NDHWGC,
             GKZYXC,

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NGCDHW,
+                                                           GKZYXC,
+                                                           NGKDHW,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_instances<
+            3,
+            NGCDHW,
+            GKZYXC,
+            NGKDHW,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NGCDHW,
+                                                           GKZYXC,
+                                                           NGKDHW,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instances<
+            3,
+            NGCDHW,
+            GKZYXC,
+            NGKDHW,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev5_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev5_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NGCDHW,
+                                                           GKZYXC,
+                                                           NGKDHW,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instances<
+            3,
+            NGCDHW,
+            GKZYXC,
+            NGKDHW,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v5>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NGCDHW,
+                                                           GKZYXC,
+                                                           NGKDHW,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_instances<
+            3,
+            NGCDHW,
+            GKZYXC,
+            NGKDHW,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp → library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instance.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
 #include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
@@ -24,19 +24,21 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<3,
-                                                                    GNDHWC,
-                                                                    GKZYXC,
-                                                                    GNDHWK,
-                                                                    ConvBwdWeightDefault>{});
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_instances<
+            3,
+            GNDHWC,
+            GKZYXC,
+            GNDHWK,
+            ConvBwdWeightDefault>{});
     // 2. Filter1x1Stride1Pad0
-    add_device_operation_instances(instances,
-                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<
-                                       3,
-                                       GNDHWC,
-                                       GKZYXC,
-                                       GNDHWK,
-                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_instances<
+            3,
+            GNDHWC,
+            GKZYXC,
+            GNDHWK,
+            ConvBwdWeightFilter1x1Stride1Pad0>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
 #include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
@@ -10,13 +10,13 @@ namespace device {
 namespace instance {
 
 // Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
-void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
                                                            NDHWGC,
                                                            GKZYXC,
                                                            NDHWGK,
                                                            BF16,
-                                                           F32,
+                                                           BF16,
                                                            BF16,
                                                            PassThrough,
                                                            PassThrough,

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/CMakeLists.txt

+# ONLY XDL_KERNELS
+set(GROUPED_CONV3D_FWD_DYNAMIC_OP
+   xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
+   xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
+   xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
+   xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp)
+
+add_instance_library(device_grouped_conv3d_fwd_dynamic_op_instance ${GROUPED_CONV3D_FWD_DYNAMIC_OP})

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_dynamic_op_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
+                                                                NDHWGC,
+                                                                GKZYXC,
+                                                                ck::Tuple<>,
+                                                                NDHWGK,
+                                                                BF16,
+                                                                BF16,
+                                                                ck::Tuple<>,
+                                                                BF16,
+                                                                PassThrough,
+                                                                PassThrough,
+                                                                DynamicUnaryOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_bf16_instances<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Tuple<>,
+                                                              NDHWGK,
+                                                              ConvFwdDefault>{});
+#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
+      // long compilation time)
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_bf16_instances<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Tuple<>,
+                                                              NDHWGK,
+                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_bf16_instances<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Tuple<>,
+                                                              NDHWGK,
+                                                              ConvFwd1x1S1P0>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_dynamic_op_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
+                                                                NDHWGC,
+                                                                GKZYXC,
+                                                                ck::Tuple<>,
+                                                                NDHWGK,
+                                                                F16,
+                                                                F16,
+                                                                ck::Tuple<>,
+                                                                F16,
+                                                                PassThrough,
+                                                                PassThrough,
+                                                                DynamicUnaryOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_f16_instances<3,
+                                                             NDHWGC,
+                                                             GKZYXC,
+                                                             Tuple<>,
+                                                             NDHWGK,
+                                                             ConvFwdDefault>{});
+#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
+      // long compilation time)
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_f16_instances<3,
+                                                             NDHWGC,
+                                                             GKZYXC,
+                                                             Tuple<>,
+                                                             NDHWGK,
+                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_f16_instances<3,
+                                                             NDHWGC,
+                                                             GKZYXC,
+                                                             Tuple<>,
+                                                             NDHWGK,
+                                                             ConvFwd1x1S1P0>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_dynamic_op_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
+                                                                NDHWGC,
+                                                                GKZYXC,
+                                                                ck::Tuple<>,
+                                                                NDHWGK,
+                                                                F32,
+                                                                F32,
+                                                                ck::Tuple<>,
+                                                                F32,
+                                                                PassThrough,
+                                                                PassThrough,
+                                                                DynamicUnaryOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_f32_instances<3,
+                                                             NDHWGC,
+                                                             GKZYXC,
+                                                             Tuple<>,
+                                                             NDHWGK,
+                                                             ConvFwdDefault>{});
+#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
+      // long compilation time)
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_f32_instances<3,
+                                                             NDHWGC,
+                                                             GKZYXC,
+                                                             Tuple<>,
+                                                             NDHWGK,
+                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_f32_instances<3,
+                                                             NDHWGC,
+                                                             GKZYXC,
+                                                             Tuple<>,
+                                                             NDHWGK,
+                                                             ConvFwd1x1S1P0>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_dynamic_op/xdl/device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_dynamic_op_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_grouped_conv3d_fwd_xdl_dynamic_op_ndhwgc_gkzyxc_ndhwgk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
+                                                                NDHWGC,
+                                                                GKZYXC,
+                                                                ck::Tuple<>,
+                                                                NDHWGK,
+                                                                int8_t,
+                                                                int8_t,
+                                                                ck::Tuple<>,
+                                                                int8_t,
+                                                                PassThrough,
+                                                                PassThrough,
+                                                                DynamicUnaryOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_int8_instances<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Tuple<>,
+                                                              NDHWGK,
+                                                              ConvFwdDefault>{});
+#if 0 // Enable with dynamic op optimizations (at now generating a lot of virtual functions cause
+      // long compilation time)
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_int8_instances<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Tuple<>,
+                                                              NDHWGK,
+                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_fwd_xdl_dynamic_op_int8_instances<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Tuple<>,
+                                                              NDHWGK,
+                                                              ConvFwd1x1S1P0>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/mha/CMakeLists.txt

@@ -27,11 +27,6 @@ rocm_install(FILES ${MHA_HEADERS} DESTINATION include/ck_tile/ops)
 # headers for building lib
 file(COPY ${MHA_HEADERS} DESTINATION ${FMHA_CPP_FOLDER})
 
-# Delete the blob file if it exists to avoid append of old content.
-if(EXISTS ${FMHA_CPP_FOLDER}/blob_list.txt)
-    file(REMOVE ${FMHA_CPP_FOLDER}/blob_list.txt)
-endif()
-
 set(FMHA_KNOWN_APIS "fwd,fwd_splitkv,fwd_appendkv,bwd")
 
 # generate a list of kernels, but not actually emit files at config stage

profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp

@@ -271,10 +271,12 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
                           << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
                           << kbatch_curr << std::endl;
 
-#if defined CK_ENABLE_FP8
+#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8
                 // set softer tolerances for fp8
-                if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
-                             is_same_v<EDataType, f8_t>)
+                if constexpr((is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
+                              is_same_v<EDataType, f8_t>) ||
+                             (is_same_v<ADataType, int8_t> || is_same_v<BDataType, int8_t> ||
+                              is_same_v<EDataType, int8_t>))
                 {
                     std::string msg = "Error: Incorrect results!";
                     double rtol     = 1e-1;
@@ -286,7 +288,7 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
                 {
 #endif
                     pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
-#if defined CK_ENABLE_FP8
+#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8
                 }
 #endif
 

profiler/include/profiler/profile_gemm_universal_batched_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CElementOp,
+          typename DeviceOp>
+bool profile_gemm_universal_batched_impl(int do_verification,
+                                         int init_method,
+                                         bool do_log,
+                                         bool time_kernel,
+                                         int M,
+                                         int N,
+                                         int K,
+                                         int BatchStrideA,
+                                         int BatchStrideB,
+                                         int BatchStrideC,
+                                         int StrideA,
+                                         int StrideB,
+                                         int StrideC,
+                                         int BatchCount,
+                                         int n_warmup,
+                                         int n_iter,
+                                         uint64_t rotating = 0)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB, BatchStrideB, BLayout{}));
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
+
+    int total_gemm_needed =
+        a_g_m_k.GetElementSpaceSizeInBytes() + b_g_k_n.GetElementSpaceSizeInBytes();
+    int rotating_count = std::max(
+        1,
+        std::min(n_iter,
+                 static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+    std::cout << "rotating count: " << rotating_count << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    if(do_verification)
+    {
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             float,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CElementOp>;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+    c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        std::unique_ptr<tensor_operation::device::BaseArgument> argument_ptr;
+        // false branch for multi d dl kernel
+
+        argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        {},
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        BatchCount,
+                                        StrideA,
+                                        StrideB,
+                                        {},
+                                        StrideC,
+                                        BatchStrideA,
+                                        BatchStrideB,
+                                        {},
+                                        BatchStrideC,
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time = invoker_ptr->Run(
+                argument_ptr.get(),
+                StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter, true, rotating_count});
+
+            std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
+
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                     sizeof(CDataType) * M * N) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " B = " << BatchCount << " M = " << M << " N = " << N << " K = " << K
+              << " StrideA = " << StrideA << " StrideB = " << StrideB << " StrideC = " << StrideC
+              << ": " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
+              << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_pool3d_fwd_impl.hpp

@@ -102,11 +102,22 @@ bool profile_pool3d_fwd_impl(PoolFwdInputParams& in_params, PoolFwdKernelParams&
     Tensor<IndexDataType> out_indices_n_c_do_ho_wo_device(
         f_host_tensor_descriptor(N, C, Do, Ho, Wo));
 
+    constexpr int inDataRangeTensor1{1};
+    constexpr int inDataRangeTensor2{5};
+    constexpr double inDataRangeTensor3{0.5};
+
     switch(in_params.init_method)
     {
-    case 0: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
-    case 1: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
-    default: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    case 0:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{inDataRangeTensor1});
+        break;
+    case 1:
+        in_n_c_di_hi_wi.GenerateTensorValue(
+            GeneratorTensor_2<InDataType>{-inDataRangeTensor2, inDataRangeTensor2});
+        break;
+    default:
+        in_n_c_di_hi_wi.GenerateTensorValue(
+            GeneratorTensor_3<InDataType>{-inDataRangeTensor3, inDataRangeTensor3});
     }
 
     DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_di_hi_wi.mDesc.GetElementSpaceSize());
@@ -229,12 +240,39 @@ bool profile_pool3d_fwd_impl(PoolFwdInputParams& in_params, PoolFwdKernelParams&
         {
             out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());
 
-            auto tolerance = 1e-3;
-            bool pass      = ck::utils::check_err(out_n_c_do_ho_wo_device.mData,
+            auto number_of_accumulations = 1;
+            static_assert(
+                ReduceOpId == ck::ReduceTensorOp::AVG || ReduceOpId == ck::ReduceTensorOp::MAX,
+                "Warning: Unhandled ReduceOpId for setting up the number of accumulations!");
+
+            if constexpr(ReduceOpId == ck::ReduceTensorOp::AVG)
+            {
+                for(size_t i = 0; i < kernel_params.window_spatial_lengths.size(); ++i)
+                {
+                    number_of_accumulations *= kernel_params.window_spatial_lengths.at(i);
+                }
+            }
+
+            auto absolute_error_threshold = 1.0;
+            switch(in_params.init_method)
+            {
+            case 0: absolute_error_threshold = static_cast<double>(inDataRangeTensor1); break;
+            case 1: absolute_error_threshold = static_cast<double>(inDataRangeTensor2); break;
+            default: absolute_error_threshold = inDataRangeTensor3;
+            }
+
+            absolute_error_threshold =
+                ck::utils::get_absolute_threshold<ComputeDataType, OutDataType>(
+                    absolute_error_threshold, number_of_accumulations);
+            auto relative_error_threshold =
+                ck::utils::get_relative_threshold<ComputeDataType, OutDataType>(
+                    number_of_accumulations);
+
+            bool pass = ck::utils::check_err(out_n_c_do_ho_wo_device.mData,
                                              out_n_c_do_ho_wo_host.mData,
                                              "Error: Incorrect results",
-                                             tolerance,
-                                             tolerance);
+                                             relative_error_threshold,
+                                             absolute_error_threshold);
 
             if constexpr(OutputIndex)
             {

profiler/src/CMakeLists.txt

@@ -59,6 +59,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
   list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
@@ -141,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   endif()
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)

profiler/src/profile_gemm_multiply_multiply.cpp

@@ -27,6 +27,7 @@ enum struct GemmDataType
     F16_F8_F16,     // 5
     F16_F16_F16_F8, // 6
     F8_F8_BF16,     // 7
+    INT8_INT8_BF16, // 8
 };
 
 #define OP_NAME "gemm_multiply_multiply"
@@ -39,7 +40,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
         printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
                "f16->f8; 7: f8->bf16, "
-               "comp f8)\n");
+               "comp f8; 8: int8->bf16)\n");
         printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
         printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
         printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
@@ -89,6 +90,8 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
     using F32  = float;
     using BF16 = ck::bhalf_t;
     using F8   = ck::f8_t;
+    using I8   = int8_t;
+    using I32  = int;
 
     using Row = ck::tensor_layout::gemm::RowMajor;
     using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -162,6 +165,11 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
         return profile(
             F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
     }
+    else if(data_type == GemmDataType::INT8_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(
+            I8{}, I8{}, I8{}, I32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
+    }
     else
     {
         std::cout << "this data_type & layout is not implemented" << std::endl;