Merge branch 'develop' into ck_tile/gemm_debug_alias

include/ck/utility/type_convert.hpp

@@ -9,7 +9,7 @@
 #include "ck/utility/array.hpp"
 
 namespace ck {
-// Define the common macro for gfx94x models
+// Define the common macro for MI300 models
 #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
 #define __gfx94__
 #endif
@@ -100,6 +100,18 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
     return type_convert<bhalf_t>(x_fp32);
 }
 
+template <>
+inline __host__ __device__ constexpr f8_ocp_t type_convert<f8_ocp_t, int>(int x)
+{
+    return f8_ocp_t{type_convert<f8_ocp_t::data_type>(x)};
+}
+
+template <>
+inline __host__ __device__ constexpr bf8_ocp_t type_convert<bf8_ocp_t, int>(int x)
+{
+    return bf8_ocp_t{type_convert<bf8_ocp_t::data_type>(x)};
+}
+
 // Convert X to Y
 template <typename Y, typename X>
 __host__ __device__ constexpr Y type_convert_sp(X x)
@@ -163,7 +175,7 @@ __host__ __device__ constexpr Y f8_convert_sr(X x);
 
 // convert fp32 to fp8 with stochastic rounding
 template <>
-inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
+inline __host__ __device__ f8_fnuz_t f8_convert_sr<f8_fnuz_t, float>(float x)
 {
     constexpr int seed = 1254739;
     uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
@@ -189,33 +201,35 @@ inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
     return utils::
-        cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
-                                                                                               rng);
+        cast_to_f8<float, f8_fnuz_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+            x, rng);
 #endif
 }
 
 // convert fp16 to fp8 with stochastic rounding
 template <>
-inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
+inline __host__ __device__ f8_fnuz_t f8_convert_sr<f8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_sr<f8_t>(type_convert<float>(x));
+    return f8_convert_sr<f8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
     constexpr int seed               = 1254739;
     uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             f8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp32 to bf8 with stochastic rounding
 template <>
-inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_sr<bf8_fnuz_t, float>(float x)
 {
     constexpr int seed = 1254739;
     uint32_t rng       = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
@@ -240,28 +254,32 @@ inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
-    return utils::
-        cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<float,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp16 to bf8 with stochastic rounding
 template <>
-inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_sr<bf8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_sr<bf8_t>(type_convert<float>(x));
+    return f8_convert_sr<bf8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
     constexpr int seed               = 1254739;
     uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
-    return utils::
-        cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
@@ -271,7 +289,7 @@ __host__ __device__ constexpr Y f8_convert_rne(X x);
 
 // convert fp32 to fp8 with rounding to nearest even
 template <>
-inline __host__ __device__ f8_t f8_convert_rne<f8_t, float>(float x)
+inline __host__ __device__ f8_fnuz_t f8_convert_rne<f8_fnuz_t, float>(float x)
 {
 #if defined(__gfx94__)
     union
@@ -296,32 +314,34 @@ inline __host__ __device__ f8_t f8_convert_rne<f8_t, float>(float x)
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
     return utils::
-        cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
-                                                                                               rng);
+        cast_to_f8<float, f8_fnuz_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+            x, rng);
 #endif
 }
 
 // convert fp16 to fp8 with rounding to nearest even
 template <>
-inline __host__ __device__ f8_t f8_convert_rne<f8_t, half_t>(half_t x)
+inline __host__ __device__ f8_fnuz_t f8_convert_rne<f8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_rne<f8_t>(type_convert<float>(x));
+    return f8_convert_rne<f8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
-    return utils::
-        cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             f8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp32 to bf8 with rounding to nearest even
 template <>
-inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, float>(float x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_rne<bf8_fnuz_t, float>(float x)
 {
 #if defined(__gfx94__)
     union
@@ -345,44 +365,59 @@ inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, float>(float x)
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
-    return utils::
-        cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<float,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
 #endif
 }
 
 // convert fp16 to bf8 with rounding to nearest even
 template <>
-inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, half_t>(half_t x)
+inline __host__ __device__ bf8_fnuz_t f8_convert_rne<bf8_fnuz_t, half_t>(half_t x)
 {
 #if defined(__gfx94__)
     // convert to float and use native converion
-    return f8_convert_rne<bf8_t>(type_convert<float>(x));
+    return f8_convert_rne<bf8_fnuz_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
     constexpr bool clip              = true;
     constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
     constexpr uint32_t rng           = 0;
-    return utils::
-        cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
-            x, rng);
+    return utils::cast_to_f8<half_t,
+                             bf8_fnuz_t,
+                             negative_zero_nan,
+                             clip,
+                             (rm == f8_rounding_mode::stochastic)>(x, rng);
+#endif
+}
+
+// convert fp32 to fp8
+template <>
+inline __host__ __device__ f8_fnuz_t type_convert<f8_fnuz_t, float>(float x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<f8_fnuz_t>(x);
+#else
+    return f8_convert_rne<f8_fnuz_t>(x);
 #endif
 }
 
 // convert fp32 to fp8
 template <>
-inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
+inline __host__ __device__ f8_ocp_t type_convert<f8_ocp_t, float>(float x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<f8_t>(x);
+    return f8_convert_sr<f8_ocp_t>(x);
 #else
-    return f8_convert_rne<f8_t>(x);
+    return f8_convert_rne<f8_ocp_t>(x);
 #endif
 }
 
 // convert fp8 to fp32
 template <>
-inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
+inline __host__ __device__ float type_convert<float, f8_fnuz_t>(f8_fnuz_t x)
 {
 #if defined(__gfx94__)
     float fval;
@@ -392,30 +427,44 @@ inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
     return fval;
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<f8_t, float, negative_zero_nan>(x);
+    return utils::cast_from_f8<f8_fnuz_t, float, negative_zero_nan>(x);
 #endif
 }
 
 template <>
-inline __host__ __device__ float2_t type_convert<float2_t, f8x2_t>(f8x2_t x)
+inline __host__ __device__ float2_t type_convert<float2_t, f8x2_fnuz_t>(f8x2_fnuz_t x)
 {
 #if defined(__gfx94__)
     const auto i16val = bit_cast<uint16_t>(x);
     return __builtin_amdgcn_cvt_pk_f32_fp8(i16val, 0);
 #else
     constexpr bool negative_zero_nan = true;
-    const auto f8x2_v                = vector_type<f8_t, 2>(x);
+    const auto f8x2_v                = vector_type<f8_fnuz_t, 2>(x);
     vector_type<float, 2> f32x2_v;
     f32x2_v.template AsType<float>()(Number<0>{}) =
-        utils::cast_from_f8<f8_t, float, negative_zero_nan>(
-            f8x2_v.template AsType<f8_t>()[Number<0>{}]);
+        utils::cast_from_f8<f8_fnuz_t, float, negative_zero_nan>(
+            f8x2_v.template AsType<f8_fnuz_t>()[Number<0>{}]);
     f32x2_v.template AsType<float>()(Number<1>{}) =
-        utils::cast_from_f8<f8_t, float, negative_zero_nan>(
-            f8x2_v.template AsType<f8_t>()[Number<1>{}]);
+        utils::cast_from_f8<f8_fnuz_t, float, negative_zero_nan>(
+            f8x2_v.template AsType<f8_fnuz_t>()[Number<1>{}]);
     return f32x2_v.template AsType<float2_t>()[Number<0>{}];
 #endif
 }
 
+template <>
+inline __host__ __device__ float2_t type_convert<float2_t, f8x2_ocp_t>(f8x2_ocp_t x)
+{
+#if CK_OCP_FP8_CVT_FAST_PATH
+    return fp8_impl::cast_to_f32x2_from_f8x2<f8_ocp_t::default_interpret>(
+        x.AsType<fp8_impl::fp8x2_storage_t>()[Number<0>{}]);
+#else
+    return float2_t{fp8_impl::cast_from_f8<float, f8_ocp_t::wm, f8_ocp_t::we, false>(
+                        x.AsType<fp8_storage_t>()[Number<0>{}]),
+                    fp8_impl::cast_from_f8<float, f8_ocp_t::wm, f8_ocp_t::we, false>(
+                        x.AsType<fp8_storage_t>()[Number<1>{}])};
+#endif
+}
+
 template <>
 inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
 {
@@ -428,42 +477,64 @@ inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
 
 // convert fp16 to fp8
 template <>
-inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
+inline __host__ __device__ f8_fnuz_t type_convert<f8_fnuz_t, half_t>(half_t x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<f8_t>(x);
+    return f8_convert_sr<f8_fnuz_t>(x);
 #else
-    return f8_convert_rne<f8_t>(x);
+    return f8_convert_rne<f8_fnuz_t>(x);
+#endif
+}
+
+// convert fp16 to fp8
+template <>
+inline __host__ __device__ f8_ocp_t type_convert<f8_ocp_t, half_t>(half_t x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<f8_ocp_t>(x);
+#else
+    return f8_convert_rne<f8_ocp_t>(x);
 #endif
 }
 
 // convert fp8 to fp16
 template <>
-inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
+inline __host__ __device__ half_t type_convert<half_t, f8_fnuz_t>(f8_fnuz_t x)
 {
 #if defined(__gfx94__)
     // use native conversion to float and convert to fp16
     return type_convert<half_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<f8_t, half_t, negative_zero_nan>(x);
+    return utils::cast_from_f8<f8_fnuz_t, half_t, negative_zero_nan>(x);
+#endif
+}
+
+// convert fp32 to bf8
+template <>
+inline __host__ __device__ bf8_fnuz_t type_convert<bf8_fnuz_t, float>(float x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<bf8_fnuz_t>(x);
+#else
+    return f8_convert_rne<bf8_fnuz_t>(x);
 #endif
 }
 
 // convert fp32 to bf8
 template <>
-inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
+inline __host__ __device__ bf8_ocp_t type_convert<bf8_ocp_t, float>(float x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<bf8_t>(x);
+    return f8_convert_sr<bf8_ocp_t>(x);
 #else
-    return f8_convert_rne<bf8_t>(x);
+    return f8_convert_rne<bf8_ocp_t>(x);
 #endif
 }
 
 // convert bf8 to fp32
 template <>
-inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
+inline __host__ __device__ float type_convert<float, bf8_fnuz_t>(bf8_fnuz_t x)
 {
 #if defined(__gfx94__)
     float fval;
@@ -473,31 +544,42 @@ inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
     return fval;
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<bf8_t, float, negative_zero_nan>(x);
+    return utils::cast_from_f8<bf8_fnuz_t, float, negative_zero_nan>(x);
+#endif
+}
+
+// convert fp16 to bf8
+template <>
+inline __host__ __device__ bf8_fnuz_t type_convert<bf8_fnuz_t, half_t>(half_t x)
+{
+#if CK_USE_SR_F8_CONVERSION
+    return f8_convert_sr<bf8_fnuz_t>(x);
+#else
+    return f8_convert_rne<bf8_fnuz_t>(x);
 #endif
 }
 
 // convert fp16 to bf8
 template <>
-inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x)
+inline __host__ __device__ bf8_ocp_t type_convert<bf8_ocp_t, half_t>(half_t x)
 {
 #if CK_USE_SR_F8_CONVERSION
-    return f8_convert_sr<bf8_t>(x);
+    return f8_convert_sr<bf8_ocp_t>(x);
 #else
-    return f8_convert_rne<bf8_t>(x);
+    return f8_convert_rne<bf8_ocp_t>(x);
 #endif
 }
 
 // convert bf8 to fp16
 template <>
-inline __host__ __device__ half_t type_convert<half_t, bf8_t>(bf8_t x)
+inline __host__ __device__ half_t type_convert<half_t, bf8_fnuz_t>(bf8_fnuz_t x)
 {
 #if defined(__gfx94__)
     // use native conversion to float and convert to fp16
     return type_convert<half_t>(type_convert<float>(x));
 #else
     constexpr bool negative_zero_nan = true;
-    return utils::cast_from_f8<bf8_t, half_t, negative_zero_nan>(x);
+    return utils::cast_from_f8<bf8_fnuz_t, half_t, negative_zero_nan>(x);
 #endif
 }
 

include/ck_tile/README.md

-# ck_tile
+[Back to the main page](../../README.md)
+# Composable Kernel Tile
 ## concept
 `ck_tile` provides a programming model with templated abstractions to enable users to implement performance-critical kernels for machine learning workloads. introduces following basic concepts to help users building your own operator
  - tensor coordinate transformation, this is the core concept of layout/index transform abstraction in both compiler time and run time.

include/ck_tile/core/utility/amd_address_space.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core/config.hpp"
+
+// Address Space for AMDGCN
+// https://llvm.org/docs/AMDGPUUsage.html#address-space
+
+namespace ck_tile {
+
+#define CK_CONSTANT_ADDRESS_SPACE __attribute__((address_space(4)))
+
+template <typename T>
+__device__ T* cast_pointer_to_generic_address_space(T CK_CONSTANT_ADDRESS_SPACE* p)
+{
+    // cast a pointer in "Constant" address space (4) to "Generic" address space (0)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+template <typename T>
+__host__ __device__ T CK_CONSTANT_ADDRESS_SPACE* cast_pointer_to_constant_address_space(T* p)
+{
+    // cast a pointer in "Generic" address space (0) to "Constant" address space (4)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T CK_CONSTANT_ADDRESS_SPACE*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+} // namespace ck_tile

include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp

@@ -998,14 +998,14 @@ struct FmhaFwdKernel
                 return pad_tensor_view(
                     q_dram_naive,
                     make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kSubQKHeaddim>{}),
-                    sequence<kPadSeqLenQ, kPadHeadDimQ>{});
+                    sequence<false, kPadHeadDimQ>{});
             }
             else
             {
                 return pad_tensor_view(
                     q_dram_naive,
                     make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0>{}),
-                    sequence<kPadSeqLenQ, kPadHeadDimQ>{});
+                    sequence<false, kPadHeadDimQ>{});
             }
         }();
         const auto k_dram = [&]() {
@@ -1019,7 +1019,7 @@ struct FmhaFwdKernel
             return pad_tensor_view(
                 k_dram_naive,
                 make_tuple(number<FmhaPipeline::kN0>{}, number<FmhaPipeline::kK0>{}),
-                sequence<kPadSeqLenK, kPadHeadDimQ>{});
+                sequence<false, kPadHeadDimQ>{});
         }();
         const auto v_dram = [&]() {
             if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
@@ -1041,7 +1041,7 @@ struct FmhaFwdKernel
                 return pad_tensor_view(
                     v_dram_transposed,
                     make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
-                    sequence<kPadHeadDimV, kPadSeqLenK>{});
+                    sequence<kPadHeadDimV, false>{});
             }
             else
             {
@@ -1055,7 +1055,7 @@ struct FmhaFwdKernel
                 return pad_tensor_view(
                     v_dram_naive,
                     make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
-                    sequence<kPadHeadDimV, kPadSeqLenK>{});
+                    sequence<false, kPadSeqLenK>{});
             }
         }();
 
@@ -1097,9 +1097,8 @@ struct FmhaFwdKernel
                         number<FmhaPipeline::kAlignmentBias>{},
                         number<1>{});
 
-                    return pad_tensor_view(bias_dram_naive,
-                                           bias_dram_window_lengths,
-                                           sequence<kPadSeqLenQ, kPadSeqLenK>{});
+                    return pad_tensor_view(
+                        bias_dram_naive, bias_dram_window_lengths, sequence<false, kPadSeqLenK>{});
                 }();
 
                 return make_tile_window(bias_dram, bias_dram_window_lengths, {i_m0, 0});

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp

@@ -339,7 +339,7 @@ struct FmhaFwdSplitKVCombineKernel
                 number<FmhaPipeline::kAlignmentOacc>{},
                 number<1>{});
 
-            auto o_acc_dram_view = pad_tensor_view(
+            const auto o_acc_dram_view = pad_tensor_view(
                 o_acc_dram_naive,
                 make_tuple(number<1>{}, number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
                 sequence<false, kPadSeqLenQ, kPadHeadDimV>{});

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp

@@ -623,14 +623,14 @@ struct FmhaFwdSplitKVKernel
                 return pad_tensor_view(
                     q_dram_naive,
                     make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kSubQKHeaddim>{}),
-                    sequence<kPadSeqLenQ, kPadHeadDimQ>{});
+                    sequence<false, kPadHeadDimQ>{});
             }
             else
             {
                 return pad_tensor_view(
                     q_dram_naive,
                     make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kK0>{}),
-                    sequence<kPadSeqLenQ, kPadHeadDimQ>{});
+                    sequence<false, kPadHeadDimQ>{});
             }
         }();
 
@@ -645,7 +645,7 @@ struct FmhaFwdSplitKVKernel
             return pad_tensor_view(
                 k_dram_naive,
                 make_tuple(number<FmhaPipeline::kN0>{}, number<FmhaPipeline::kK0>{}),
-                sequence<kPadSeqLenK, kPadHeadDimQ>{});
+                sequence<false, kPadHeadDimQ>{});
         };
         const auto k_dram = [&]() {
             if constexpr(kIsPagedKV)
@@ -678,7 +678,7 @@ struct FmhaFwdSplitKVKernel
                 return pad_tensor_view(
                     v_dram_transposed,
                     make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
-                    sequence<kPadHeadDimV, kPadSeqLenK>{});
+                    sequence<kPadHeadDimV, false>{});
             }
             else
             {
@@ -692,7 +692,7 @@ struct FmhaFwdSplitKVKernel
                 return pad_tensor_view(
                     v_dram_naive,
                     make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
-                    sequence<kPadHeadDimV, kPadSeqLenK>{});
+                    sequence<false, kPadSeqLenK>{});
             }
         };
         const auto v_dram = [&]() {
@@ -804,9 +804,8 @@ struct FmhaFwdSplitKVKernel
                         number<FmhaPipeline::kAlignmentBias>{},
                         number<1>{});
 
-                    return pad_tensor_view(bias_dram_naive,
-                                           bias_dram_window_lengths,
-                                           sequence<kPadSeqLenQ, kPadSeqLenK>{});
+                    return pad_tensor_view(
+                        bias_dram_naive, bias_dram_window_lengths, sequence<false, kPadSeqLenK>{});
                 }();
 
                 return make_tile_window(bias_dram, bias_dram_window_lengths, {i_m0, 0});

include/ck_tile/ops/gemm.hpp

@@ -27,6 +27,7 @@
 #include "ck_tile/ops/gemm/block/block_universal_gemm_as_bs_cr.hpp"
 #include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
 #include "ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp"
+#include "ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp"
 #include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp"

include/ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp

@@ -35,4 +35,40 @@ struct GemmTilePartitioner
         return make_tuple(iM, iN);
     }
 };
+
+template <typename BlockGemmShape_>
+struct GemmTile1DPartitioner
+{
+    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
+
+    static constexpr index_t MPerBlock = BlockGemmShape::kM;
+    static constexpr index_t NPerBlock = BlockGemmShape::kN;
+    static constexpr index_t KPerBlock = BlockGemmShape::kK;
+
+    CK_TILE_HOST static constexpr auto GridSize(index_t M, index_t N)
+    {
+        index_t GridDimX = (M + MPerBlock - 1) / MPerBlock;
+        index_t GridDimY = (N + NPerBlock - 1) / NPerBlock;
+        return dim3(GridDimX * GridDimY, 1, 1);
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr auto GetNBlock(index_t N)
+    {
+        return integer_divide_ceil(N, NPerBlock);
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr auto GetLoopNum(index_t K)
+    {
+        return integer_divide_ceil(K, KPerBlock);
+    }
+
+    CK_TILE_DEVICE auto operator()(index_t blockOffset, index_t NBlockSize)
+    {
+        index_t iM = __builtin_amdgcn_readfirstlane((blockIdx.x - blockOffset) /
+                                                    GetNBlock(NBlockSize) * MPerBlock);
+        index_t iN = __builtin_amdgcn_readfirstlane((blockIdx.x - blockOffset) %
+                                                    GetNBlock(NBlockSize) * NPerBlock);
+        return make_tuple(iM, iN);
+    }
+};
 } // namespace ck_tile

include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <string>
+
+#include "ck_tile/core/numeric/math.hpp"
+#include "ck_tile/core/utility/literals.hpp"
+#include "ck_tile/core/utility/amd_address_space.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/common.hpp"
+#include "ck_tile/host.hpp"
+
+namespace ck_tile {
+
+struct GroupedGemmHostArgs
+{
+    const void* a_ptr;
+    const void* b_ptr;
+    void* c_ptr;
+    index_t M;
+    index_t N;
+    index_t K;
+    index_t stride_A;
+    index_t stride_B;
+    index_t stride_C;
+};
+
+template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
+struct GroupedGemmKernel
+{
+    using TilePartitioner                    = remove_cvref_t<TilePartitioner_>;
+    using GemmPipeline                       = remove_cvref_t<GemmPipeline_>;
+    using EpiloguePipeline                   = remove_cvref_t<EpiloguePipeline_>;
+    using ALayout                            = remove_cvref_t<typename GemmPipeline::ALayout>;
+    using BLayout                            = remove_cvref_t<typename GemmPipeline::BLayout>;
+    using CLayout                            = remove_cvref_t<typename GemmPipeline::CLayout>;
+    static constexpr index_t KernelBlockSize = GemmPipeline::BlockSize;
+
+    using ADataType = remove_cvref_t<typename GemmPipeline::ADataType>;
+    using BDataType = remove_cvref_t<typename GemmPipeline::BDataType>;
+    using CDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
+
+    struct GemmTransKernelArg
+    {
+        GroupedGemmHostArgs group_karg;
+        ck_tile::index_t block_start;
+        ck_tile::index_t block_end;
+
+        GemmTransKernelArg() = default;
+        GemmTransKernelArg(GroupedGemmHostArgs&& karg, index_t bl_start, index_t bl_end)
+            : group_karg{karg}, block_start{bl_start}, block_end{bl_end}
+        {
+        }
+    };
+
+    __host__ static size_t GetWorkSpaceSize(const std::vector<GroupedGemmHostArgs>& gemm_descs)
+    {
+        return gemm_descs.size() * sizeof(GemmTransKernelArg);
+    }
+
+    __host__ static constexpr auto BlockSize() { return dim3(KernelBlockSize); }
+
+    using Hargs = GroupedGemmHostArgs;
+
+    __host__ static constexpr auto GridSize(const std::vector<Hargs>& gemm_descs)
+    {
+        index_t grid_size = 0;
+        for(const auto& it_desc : gemm_descs)
+        {
+            const auto dim3 = TilePartitioner::GridSize(it_desc.M, it_desc.N);
+            grid_size += dim3.x * dim3.y * 1;
+        }
+        return dim3(grid_size, 1, 1);
+    }
+
+    CK_TILE_HOST static auto MakeKargs(const std::vector<Hargs>& gemm_descs)
+    {
+        std::vector<GemmTransKernelArg> gemm_kernel_args_;
+        index_t group_count = ck_tile::type_convert<ck_tile::index_t>(gemm_descs.size());
+        index_t grid_size   = 0;
+        gemm_kernel_args_.reserve(group_count);
+
+        for(std::size_t i = 0; i < gemm_descs.size(); ++i)
+        {
+            const index_t M = gemm_descs[i].M;
+            const index_t N = gemm_descs[i].N;
+            const index_t K = gemm_descs[i].K;
+
+            if(M == 0 || N == 0 || K == 0)
+            {
+                continue;
+            }
+
+            const index_t stride_a = gemm_descs[i].stride_A;
+            const index_t stride_b = gemm_descs[i].stride_B;
+            const index_t stride_c = gemm_descs[i].stride_C;
+
+            const auto dim3             = TilePartitioner::GridSize(M, N);
+            const index_t grid_size_grp = dim3.x * 1 * 1;
+
+            const index_t block_start = grid_size;
+            const index_t block_end   = grid_size + grid_size_grp;
+
+            grid_size += grid_size_grp;
+
+            auto karg = GroupedGemmHostArgs{type_convert<const ADataType*>(gemm_descs[i].a_ptr),
+                                            type_convert<const BDataType*>(gemm_descs[i].b_ptr),
+                                            type_convert<CDataType*>(gemm_descs[i].c_ptr),
+                                            M,
+                                            N,
+                                            K,
+                                            stride_a,
+                                            stride_b,
+                                            stride_c};
+
+            gemm_kernel_args_.emplace_back(std::move(karg), block_start, block_end);
+        }
+
+        return gemm_kernel_args_;
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
+    }
+
+    CK_TILE_DEVICE void Run(const Hargs& kargs, const index_t block_start) const
+    {
+        const auto [i_m, i_n] = TilePartitioner{}(block_start, kargs.N);
+        // options
+        const ADataType* a_start = static_cast<const ADataType*>(kargs.a_ptr);
+        const BDataType* b_start = static_cast<const BDataType*>(kargs.b_ptr);
+        // Convert pointers to tensor views
+        auto a_tensor_view = [&]() {
+            if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    a_start,
+                    make_tuple(kargs.M, kargs.K),
+                    make_tuple(kargs.stride_A, 1),
+                    number<GemmPipeline::VectorSizeA>{},
+                    number<1>{});
+            }
+            else
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    a_start,
+                    make_tuple(kargs.M, kargs.K),
+                    make_tuple(1, kargs.stride_A),
+                    number<1>{},
+                    number<1>{});
+            }
+        }();
+
+        auto b_tensor_view = [&]() {
+            if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    b_start,
+                    make_tuple(kargs.N, kargs.K),
+                    make_tuple(1, kargs.stride_B),
+                    number<1>{},
+                    number<1>{});
+            }
+            else
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    b_start,
+                    make_tuple(kargs.N, kargs.K),
+                    make_tuple(kargs.stride_B, 1),
+                    number<GemmPipeline::VectorSizeB>{},
+                    number<1>{});
+            }
+        }();
+
+        auto a_pad_view = [&]() {
+            if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
+            {
+                return pad_tensor_view(a_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<false, GemmPipeline::kPadK>{});
+            }
+            else
+            {
+                return pad_tensor_view(a_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<GemmPipeline::kPadM, false>{});
+            }
+        }();
+        // clang-format on
+
+        auto a_block_window = make_tile_window(
+            a_pad_view,
+            make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::KPerBlock>{}),
+            {i_m, 0});
+
+        auto b_pad_view = [&]() {
+            if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
+            {
+                return pad_tensor_view(b_tensor_view,
+                                       make_tuple(number<TilePartitioner::NPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<false, GemmPipeline::kPadK>{});
+            }
+            else
+            {
+                return pad_tensor_view(b_tensor_view,
+                                       make_tuple(number<TilePartitioner::NPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<GemmPipeline::kPadN, false>{});
+            }
+        }();
+
+        auto b_block_window = make_tile_window(
+            b_pad_view,
+            make_tuple(number<TilePartitioner::NPerBlock>{}, number<TilePartitioner::KPerBlock>{}),
+            {i_n, 0});
+
+        // allocate LDS
+        __shared__ char smem_ptr[GetSmemSize()];
+
+        const index_t num_loop = TilePartitioner::GetLoopNum(kargs.K);
+
+        // Run GEMM cooperatively by whole wokrgroup.
+        auto c_block_tile =
+            GemmPipeline{}.template operator()(a_block_window, b_block_window, num_loop, smem_ptr);
+
+        CDataType* c_start = static_cast<CDataType*>(kargs.c_ptr);
+        auto c_tensor_view = [&]() {
+            if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    c_start,
+                    make_tuple(kargs.M, kargs.N),
+                    make_tuple(kargs.stride_C, 1),
+                    number<GemmPipeline::VectorSizeC>{},
+                    number<1>{});
+            }
+            else
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    c_start,
+                    make_tuple(kargs.M, kargs.N),
+                    make_tuple(1, kargs.stride_C),
+                    number<1>{},
+                    number<1>{});
+            }
+        }();
+
+        auto c_pad_view = [&]() {
+            if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
+            {
+                return pad_tensor_view(c_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::NPerBlock>{}),
+                                       sequence<false, GemmPipeline::kPadN>{});
+            }
+            else
+            {
+                return pad_tensor_view(c_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::NPerBlock>{}),
+                                       sequence<GemmPipeline::kPadM, false>{});
+            }
+        }();
+        auto CBlockWindow_pad = make_tile_window(
+            c_pad_view,
+            make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::NPerBlock>{}),
+            {i_m, i_n});
+
+        EpiloguePipeline{}(CBlockWindow_pad, c_block_tile);
+    }
+
+    CK_TILE_DEVICE void operator()(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                   int group_count) const
+    {
+        const index_t block_id   = ck_tile::get_block_1d_id();
+        const auto gemm_desc_ptr = reinterpret_cast<const GemmTransKernelArg*>(
+            cast_pointer_to_generic_address_space(gemm_descs_const));
+
+        index_t left     = 0;
+        index_t right    = group_count;
+        index_t group_id = index_t((left + right) / 2);
+
+        while((!(block_id >= gemm_desc_ptr[group_id].block_start &&
+                 block_id < gemm_desc_ptr[group_id].block_end)) &&
+              left <= right)
+        {
+            if(block_id < gemm_desc_ptr[group_id].block_start)
+            {
+                right = group_id;
+            }
+            else
+            {
+                left = group_id;
+            }
+            group_id = index_t((left + right) / 2);
+        }
+
+        Run(gemm_desc_ptr[group_id].group_karg, gemm_desc_ptr[group_id].block_start);
+    }
+};
+
+} // namespace ck_tile

library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp

 // 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.
 
 #pragma once
 
@@ -62,9 +62,9 @@ struct ReferenceGemm : public device::BaseOperator
             auto f_mk_kn_mn = [&](auto m, auto n) {
                 const int K = arg.a_m_k_.mDesc.GetLengths()[1];
 
-                AccDataType v_acc = 0;
-                ComputeTypeA v_a  = 0;
-                ComputeTypeB v_b  = 0;
+                AccDataType v_acc{0};
+                ComputeTypeA v_a{0};
+                ComputeTypeB v_b{0};
 
                 for(int k = 0; k < K; ++k)
                 {
@@ -93,7 +93,7 @@ struct ReferenceGemm : public device::BaseOperator
                         ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
                 }
 
-                CDataType v_c = 0;
+                CDataType v_c{0};
 
                 arg.c_element_op_(v_c, v_acc);
 

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -62,7 +62,7 @@ function(add_instance_library INSTANCE_NAME)
     endforeach()
     # Do not build mha instances if gfx94 or gfx90a targets are not on the target list
     foreach(source IN LISTS ARGN)
-    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
+	    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
          message("removing mha instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
@@ -346,7 +346,7 @@ if(CK_DEVICE_CONV_INSTANCES)
 endif()
 if(CK_DEVICE_MHA_INSTANCES)
         set(gpu_list ${INST_TARGETS})
-        if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
+	if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
             add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
             add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
             target_compile_features(device_mha_operations PUBLIC)

library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f8_instance.cpp

@@ -15,7 +15,7 @@ void add_device_pool3d_fwd_ndhwc_f8_instances(
         instances)
 {
     add_device_operation_instances(
-        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F8, ReduceOpId, false>{});
+        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F32, ReduceOpId, false>{});
 }
 
 void add_device_pool3d_fwd_ndhwc_index_f8_instances(
@@ -23,7 +23,7 @@ void add_device_pool3d_fwd_ndhwc_index_f8_instances(
         instances)
 {
     add_device_operation_instances(
-        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F8, ReduceOpId, true>{});
+        instances, device_pool3d_fwd_ndhwc_instances<F8, F8, I32, F32, ReduceOpId, true>{});
 }
 
 } // namespace instance

profiler/README.md

+[Back to the main page](../README.md)
+# Composable Kernel profiler
 ## Profile GEMM kernels
 ```bash
 #arg1: tensor operation (gemm=GEMM)
@@ -180,3 +182,13 @@ Note: Column to image kernel adds to the output memory, this will cause output b
 ################            op datatype  verify  init  log  time  dim0 dim1 dim2 in_stride0 in_stride1 in_stride2 out_stride0 out_stride1 out_stride2
 ./bin/ckProfiler permute_scale        0       1     1    0     1    64   64   64       4096         64          1           1          64        4096
 ```
+
+## Convert MIOpen driver command to CKProfiler
+
+```bash
+python3 ../script/convert_miopen_driver_to_profiler.py
+/opt/rocm/bin/MIOpenDriver conv -n 32 -c 64 -H 28 -W 28 -k 64 -y 3 -x 3
+-p 1 -q 1 -u 2 -v 2 -l 1 -j 1 -m conv -g 32 -F 1 -t 1 
+```
+
+Only convolution driver is supported.

profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp

 // 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.
 
 #pragma once
 
@@ -150,7 +150,7 @@ bool profile_batched_gemm_bias_softmax_gemm_permute_impl(bool do_verification,
         break;
     default:
         a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
         d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
     }

profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp

 // 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.
 
 #pragma once
 
@@ -157,7 +157,7 @@ bool profile_batched_gemm_gemm_impl(bool do_verification,
         break;
     default:
         a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
     }
 

profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp

 // 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.
 
 #pragma once
 
@@ -174,7 +174,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
         break;
     default:
         a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
     }
 

profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp

 // 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.
 
 #pragma once
 
@@ -140,7 +140,7 @@ bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
         break;
     default:
         a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<B0DataType, 1>{});
         b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
     }
 

profiler/include/profiler/profile_gemm_impl.hpp

 // 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.
 
 #pragma once
 
@@ -74,8 +74,8 @@ int profile_gemm_impl(int do_verification,
     switch(init_method)
     {
     case 0:
-        ck::utils::FillConstant<ADataType>{static_cast<ADataType>(1.f)}(a_m_k);
-        ck::utils::FillConstant<BDataType>{static_cast<BDataType>(1.f)}(b_k_n);
+        ck::utils::FillConstant<ADataType>{type_convert<ADataType>(1.f)}(a_m_k);
+        ck::utils::FillConstant<BDataType>{type_convert<BDataType>(1.f)}(b_k_n);
         break;
     case 1:
         ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);

test/ck_tile/CMakeLists.txt

 add_subdirectory(image_to_column)
 add_subdirectory(gemm)
 add_subdirectory(batched_gemm)
+add_subdirectory(grouped_gemm)

test/ck_tile/grouped_gemm/CMakeLists.txt

+# Currently ck_tile is only built on gfx9
+if(GPU_TARGETS MATCHES "gfx9")
+    add_gtest_executable(test_ck_tile_grouped_gemm test_grouped_gemm.cpp)
+endif()