Merge branch 'develop' into enable_navi_32_ci

src/targets/gpu/include/migraphx/gpu/device/argmax.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -36,7 +36,8 @@ namespace device {
 void MIGRAPHX_DEVICE_EXPORT argmax(hipStream_t stream,
                                    const argument& result,
                                    const argument& arg,
-                                   int64_t axis);
+                                   int64_t axis,
+                                   bool select_last_index);
 
 } // namespace device
 } // namespace gpu

src/targets/gpu/include/migraphx/gpu/device/argmin.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -36,7 +36,8 @@ namespace device {
 void MIGRAPHX_DEVICE_EXPORT argmin(hipStream_t stream,
                                    const argument& result,
                                    const argument& arg,
-                                   int64_t axis);
+                                   int64_t axis,
+                                   bool select_last_index);
 
 } // namespace device
 } // namespace gpu

src/targets/gpu/include/migraphx/gpu/fuse_mlir.hpp

@@ -38,6 +38,7 @@ MIGRAPHX_GPU_EXPORT bool mlir_enabled();
 struct MIGRAPHX_GPU_EXPORT fuse_mlir
 {
     context* ctx = nullptr;
+    bool enable_extra = false;
     std::string name() const { return "gpu::fuse_mlir"; }
     void apply(module_pass_manager& mpm) const;
 };

src/targets/gpu/include/migraphx/gpu/fuse_ops.hpp

@@ -24,7 +24,6 @@
 #ifndef MIGRAPHX_GUARD_RTGLIB_FUSE_OPS_HPP
 #define MIGRAPHX_GUARD_RTGLIB_FUSE_OPS_HPP
 
-#include <migraphx/config.hpp>
 #include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
@@ -34,7 +33,7 @@ struct module;
 
 namespace gpu {
 
-struct fuse_ops
+struct MIGRAPHX_GPU_EXPORT fuse_ops
 {
     context* ctx   = nullptr;
     bool fast_math = true;

src/targets/gpu/include/migraphx/gpu/gemm_softmax_gemm.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_GPU_GEMM_SOFTMAX_GEMM_HPP
+#define MIGRAPHX_GUARD_GPU_GEMM_SOFTMAX_GEMM_HPP
+
+#include <migraphx/make_op.hpp>
+#include <migraphx/check_shapes.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct gemm_softmax_gemm
+{
+    operation op = make_op("dot");
+    float scale  = 1.0;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.op, "op"), f(self.scale, "scale"));
+    }
+
+    std::string name() const { return "gpu::gemm_softmax_gemm"; }
+
+    void check_gemm_shape(const shape& s) const
+    {
+        if(not contains(range(s.strides().rbegin(), s.strides().rbegin() + 3), 1))
+            MIGRAPHX_THROW("Invalid shape for " + name());
+    }
+
+    shape compute_shape(std::vector<shape> inputs, const std::vector<module_ref>&) const
+    {
+        check_shapes{inputs, *this}.same_ndims();
+        if(inputs.size() < 3)
+            MIGRAPHX_THROW(name() + ": Expected 3 inputs but got " + to_string(inputs.size()));
+        auto a  = inputs[0];
+        auto b  = inputs[1];
+        auto b1 = inputs[2];
+        for(const auto& input : inputs)
+        {
+            check_gemm_shape(input);
+        }
+        return op.compute_shape({op.compute_shape({a, b}), b1});
+    }
+
+    static bool is_ck_supported_type(shape::type_t t) { return contains({shape::half_type}, t); }
+};
+
+} // namespace gpu
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_GPU_GEMM_SOFTMAX_GEMM_HPP

src/targets/gpu/include/migraphx/gpu/prefuse_ops.hpp

@@ -24,7 +24,7 @@
 #ifndef MIGRAPHX_GUARD_GPU_PREFUSE_OPS_HPP
 #define MIGRAPHX_GUARD_GPU_PREFUSE_OPS_HPP
 
-#include <migraphx/config.hpp>
+#include <migraphx/gpu/config.hpp>
 #include <string>
 
 namespace migraphx {
@@ -34,7 +34,7 @@ struct module_pass_manager;
 
 namespace gpu {
 
-struct prefuse_ops
+struct MIGRAPHX_GPU_EXPORT prefuse_ops
 {
     std::string name() const { return "gpu::prefuse_ops"; }
     void apply(module_pass_manager& mpm) const;

src/targets/gpu/include/migraphx/gpu/time_op.hpp

@@ -32,7 +32,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-MIGRAPHX_GPU_EXPORT std::pair<double, double>
+MIGRAPHX_GPU_EXPORT double
 time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n = 100);
 
 } // namespace gpu

src/targets/gpu/jit/ck_gemm.cpp

@@ -27,6 +27,7 @@
 #include <migraphx/make_op.hpp>
 #include <migraphx/gpu/context.hpp>
 
+#include <migraphx/gpu/ck.hpp>
 #include <migraphx/env.hpp>
 #include <migraphx/file_buffer.hpp>
 #include <migraphx/gpu/compile_gen.hpp>
@@ -37,8 +38,6 @@
 #include <migraphx/reduce_dims.hpp>
 #include <migraphx/stringutils.hpp>
 
-#include "ck/host/device_gemm_multiple_d.hpp"
-
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
@@ -46,12 +45,6 @@ namespace gpu {
 
 using namespace migraphx::gpu::gen; // NOLINT
 
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_LOG_CK_GEMM);
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_TUNING);
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_TUNING_VALUE);
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_DEBUG);
-MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TUNE_CK);
-
 // NOLINTNEXTLINE
 static const char* const ck_gemm_kernel = R"__migraphx__(
 #include <args.hpp>
@@ -79,220 +72,10 @@ MIGRAPHX_GLOBAL void ${kernel}(${params})
 
 )__migraphx__";
 
-// NOLINTNEXTLINE
-static const char* const disable_warning_pragma = R"__migraphx__(
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Weverything"
-${content}
-#pragma clang diagnostic pop
-)__migraphx__";
-
-template <class P>
-static std::string ck_disable_warnings(P p)
-{
-    return interpolate_string(disable_warning_pragma,
-                              {{"content", std::string{p.first, p.second}}});
-}
-
-static std::unordered_map<std::string, std::string> create_ck_header_strings()
-{
-    std::unordered_map<std::string, std::string> result;
-    auto ck_headers = ck::host::GetHeaders();
-
-    std::transform(
-        ck_headers.begin(), ck_headers.end(), std::inserter(result, result.begin()), [&](auto&& p) {
-            return std::make_pair(p.first, ck_disable_warnings(p.second));
-        });
-    return result;
-}
-
-static std::vector<src_file> create_ck_headers()
-{
-    static const auto& header_strings = create_ck_header_strings();
-    std::vector<src_file> srcs;
-    std::transform(
-        header_strings.begin(), header_strings.end(), std::back_inserter(srcs), [&](auto&& p) {
-            return src_file{fs::path{p.first},
-                            {p.second.data(), p.second.data() + p.second.size()}};
-        });
-    return srcs;
-}
-
-static const std::vector<src_file>& ck_headers()
-{
-    static const auto& headers = create_ck_headers();
-    return headers;
-}
-
-static bool transposed_matrix(const shape& s) { return s.strides().back() != 1; }
-
-using tuning_entry = std::pair<std::vector<shape>, size_t>;
-static std::vector<tuning_entry> read_tuning(const std::string& s)
-{
-    if(not fs::exists(s))
-        return {};
-    return from_value<std::vector<tuning_entry>>(from_json_string(read_string(s)));
-}
-
-static float matrix_distance(const shape& x, const shape& y)
-{
-    if(x.type() != y.type())
-        return std::numeric_limits<float>::max();
-    if(transposed_matrix(x) != transposed_matrix(y))
-        return std::numeric_limits<float>::max();
-    auto sum_squared = std::inner_product(x.lens().rbegin(),
-                                          x.lens().rbegin() + 2,
-                                          y.lens().rbegin(),
-                                          0,
-                                          std::plus<>{},
-                                          [](auto a, auto b) { return (a - b) * (a - b); });
-    return std::sqrt(sum_squared);
-}
-
-static std::size_t get_tuning_for(const std::vector<shape>& inputs)
-{
-    static auto tuning = read_tuning(string_value_of(MIGRAPHX_CK_TUNING{}, ""));
-    if(tuning.empty())
-    {
-        std::cout << "*********** Warning: No CK tuning! for config:" << std::endl;
-        std::cout << "  " << inputs[0] << std::endl;
-        std::cout << "  " << inputs[1] << std::endl;
-        std::cout << "  " << inputs[2] << std::endl;
-    }
-    auto it = std::find_if(
-        tuning.begin(), tuning.end(), [&](const auto& p) { return p.first == inputs; });
-    if(it == tuning.end())
-    {
-        std::cout << "*********** Warning: CK tuning missing for config!" << std::endl;
-        std::cout << "  " << inputs[0] << std::endl;
-        std::cout << "  " << inputs[1] << std::endl;
-        std::cout << "  " << inputs[2] << std::endl;
-        std::vector<std::pair<float, std::size_t>> w;
-        std::transform(tuning.begin(), tuning.end(), std::back_inserter(w), [&](const auto& p) {
-            if(inputs.size() < 3 or p.first.size() < 3)
-                MIGRAPHX_THROW("Invalid CK config");
-            auto avg_distance = std::inner_product(
-                p.first.begin(),
-                p.first.begin() + 3,
-                inputs.begin(),
-                0.0f,
-                std::plus<>{},
-                [](const auto& x, const auto& y) { return matrix_distance(x, y) / 3.0f; });
-            return std::make_pair(avg_distance, p.second);
-        });
-        std::sort(w.begin(), w.end());
-        std::size_t default_value = 4;
-        if(not w.empty())
-            default_value = w.front().second;
-        auto tuning_val = value_of(MIGRAPHX_CK_TUNING_VALUE{}, default_value);
-        std::cout << "*********** Warning: CK try tuning: " << tuning_val << std::endl;
-        return tuning_val;
-    }
-    return it->second;
-}
-
 struct ck_gemm_compiler : compiler<ck_gemm_compiler>
 {
-    static std::string get_layout(const shape& s)
-    {
-        return transposed_matrix(s) ? "ck::tensor_layout::gemm::ColumnMajor"
-                                    : "ck::tensor_layout::gemm::RowMajor";
-    }
-
-    static ck::host::DataType get_type(const shape& s)
-    {
-        if(s.type() == shape::half_type)
-            return ck::host::DataType::Half;
-        else if(s.type() == shape::float_type)
-            return ck::host::DataType::Float;
-        else if(s.type() == shape::int8_type)
-            return ck::host::DataType::Int8;
-        else if(s.type() == shape::int32_type)
-            return ck::host::DataType::Int32;
-        MIGRAPHX_THROW("Unsupported ck type");
-    }
-
-    template <class Iterator, class F>
-    static std::string ck_tuple(Iterator start, Iterator last, F f)
-    {
-        std::vector<std::string> s;
-        std::transform(start, last, std::back_inserter(s), f);
-        return "ck::Tuple<" + join_strings(s, ",") + ">";
-    }
-
-    static std::vector<shape> adjust_inputs(std::vector<shape> inputs, bool& swap_inputs)
-    {
-        swap_inputs  = false;
-        auto c_shape = inputs.back();
-        if(not transposed_matrix(c_shape))
-            return inputs;
-        std::vector<int64_t> perm(c_shape.lens().size());
-        std::iota(perm.begin(), perm.end(), 0);
-        std::swap(perm[perm.size() - 1], perm[perm.size() - 2]);
-        std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](shape s) {
-            return reorder_shape(s, perm);
-        });
-        swap_inputs = true;
-        return inputs;
-    }
-
-    static std::size_t get_batch_count(const shape& s)
-    {
-        return std::accumulate(
-            s.lens().rbegin() + 2, s.lens().rend(), std::size_t{1}, std::multiplies<std::size_t>());
-    }
-
-    static void fold_batch_dims(shape& s)
-    {
-        auto lens = s.lens();
-        if(lens.size() <= 2)
-            return;
-        auto batch_count = get_batch_count(s);
-        auto m1          = lens.at(lens.size() - 2);
-        auto m2          = lens.at(lens.size() - 1);
-        if(transposed_matrix(s))
-            s = shape{s.type(), {m1, m2 * batch_count}};
-        else
-            s = shape{s.type(), {m1 * batch_count, m2}};
-    }
-
-    static void remove_batch_dims(shape& s)
-    {
-        auto lens = s.lens();
-        if(lens.size() <= 2)
-            return;
-        auto m1 = lens.at(lens.size() - 2);
-        auto m2 = lens.at(lens.size() - 1);
-        s       = shape{s.type(), {m1, m2}};
-    }
-
     std::vector<std::string> names() const { return {"ck_gemm", "gpu::ck_gemm"}; }
 
-    static bool standard_batch(const shape& s)
-    {
-        if(s.lens().size() < 3)
-            return true;
-        std::vector<std::size_t> lens(s.lens().begin(), s.lens().end() - 2);
-        std::vector<std::size_t> strides(s.strides().begin(), s.strides().end() - 2);
-        auto base = *(s.lens().end() - 2) * *(s.lens().end() - 1);
-        std::transform(strides.begin(), strides.end(), strides.begin(), [&](auto stride) {
-            return stride / base;
-        });
-        return shape{s.type(), lens, strides}.standard();
-    }
-
-    bool can_fold_batch(const std::vector<shape>& inputs) const
-    {
-        const auto& b_shape = inputs[1];
-        if(std::any_of(inputs.begin() + 2, inputs.end() - 1, [](auto input) {
-               return not standard_batch(input);
-           }))
-            return false;
-        const auto& b_strides = b_shape.strides();
-        return std::all_of(
-            b_strides.begin(), b_strides.end() - 2, [](auto stride) { return stride == 0; });
-    }
-
     ck::host::device_gemm_multiple_d::Problem create_problem(const std::vector<shape>& inputs,
                                                              const value& v) const
     {
@@ -301,8 +84,7 @@ struct ck_gemm_compiler : compiler<ck_gemm_compiler>
         const auto& c_shape = inputs.back();
 
         // cppcheck-suppress unreadVariable
-        auto rank = a_shape.ndim();
-
+        auto rank        = a_shape.ndim();
         auto batch_count = get_batch_count(c_shape);
         auto m           = c_shape.lens()[rank - 2];
         m                = can_fold_batch(inputs) ? m * batch_count : m;
@@ -352,12 +134,8 @@ struct ck_gemm_compiler : compiler<ck_gemm_compiler>
 
     operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
     {
-        const auto& a_shape = inputs[0];
-        const auto& b_shape = inputs[1];
         const auto& c_shape = inputs.back();
-        auto tuning_value   = v.get("tuning_value", 4);
-        if(not v.contains("tuning_value"))
-            tuning_value = get_tuning_for({a_shape, b_shape, c_shape});
+        auto tuning_value   = v.get("tuning_value", 34);
         auto batch_count = get_batch_count(c_shape);
         auto problem     = create_problem(inputs, v);
 

src/targets/gpu/jit/ck_gemm_softmax_gemm.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <fstream>
+#include <migraphx/filesystem.hpp>
+#include <migraphx/gpu/compiler.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/gpu/context.hpp>
+
+#include <migraphx/env.hpp>
+#include <migraphx/file_buffer.hpp>
+#include <migraphx/gpu/ck.hpp>
+#include <migraphx/gpu/compile_gen.hpp>
+#include <migraphx/gpu/compile_hip.hpp>
+#include <migraphx/gpu/compile_hip_code_object.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/reduce_dims.hpp>
+#include <migraphx/stringutils.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+namespace gpu {
+
+using namespace migraphx::gpu::gen; // NOLINT
+
+// NOLINTNEXTLINE
+static const char* const ck_gemm_softmax_gemm_kernel = R"__migraphx__(
+#include <args.hpp>
+#include <migraphx/kernels/ck_gemm_softmax_gemm.hpp>
+#include <migraphx/kernels/pointwise.hpp>
+#include <migraphx/kernels/ops.hpp>
+#include <migraphx/kernels/integral_constant.hpp>
+#include <migraphx/kernels/generic_constant.hpp>
+#include <${include}>
+
+namespace migraphx {
+
+${preamble}
+
+extern "C" {
+
+MIGRAPHX_GLOBAL void ${kernel}(${params})
+{
+    transform_args(make_tensors(), rotate_last())(${args})([](auto... xs) {
+        auto settings = make_ck_gemm_softmax_gemm_settings(MIGRAPHX_MAKE_CONSTANT(float{SCALE}));
+        ck_gemm_softmax_gemm<${solution}, ${blocks_per_batch}>(settings, xs...);
+    });
+}
+
+}
+
+} // namespace migraphx
+
+)__migraphx__";
+
+struct ck_gemm_softmax_gemm_compiler : compiler<ck_gemm_softmax_gemm_compiler>
+{
+    std::vector<std::string> names() const
+    {
+        return {"ck_gemm_softmax_gemm", "gpu::ck_gemm_softmax_gemm"};
+    }
+
+    ck::host::device_batched_gemm_softmax_gemm::Problem
+    create_problem(const std::vector<shape>& inputs, const value&) const
+    {
+        const auto& a_shape  = inputs[0];
+        const auto& b_shape  = inputs[1];
+        const auto& b1_shape = inputs[2];
+        const auto& c_shape  = inputs.back();
+
+        // cppcheck-suppress unreadVariable
+        auto rank        = a_shape.ndim();
+        auto batch_count = get_batch_count(c_shape);
+        auto m           = c_shape.lens()[rank - 2];
+        m                = can_fold_batch(inputs) ? m * batch_count : m;
+        auto n           = c_shape.lens().back();
+        auto k           = a_shape.lens().back();
+        auto o           = c_shape.lens().back();
+
+        const bool trans_a  = transposed_matrix(a_shape);
+        const bool trans_b  = transposed_matrix(b_shape);
+        const bool trans_b1 = transposed_matrix(b1_shape);
+        const bool trans_c  = transposed_matrix(c_shape);
+        const auto a_type   = get_type(a_shape);
+        const auto b_type   = get_type(b_shape);
+        const auto b1_type  = get_type(b1_shape);
+        const auto c_type   = get_type(c_shape);
+
+        std::string ck_passthrough = "ck_passthrough";
+        return ck::host::device_batched_gemm_softmax_gemm::Problem{m,
+                                                                   n,
+                                                                   k,
+                                                                   o,
+                                                                   trans_a,
+                                                                   trans_b,
+                                                                   trans_b1,
+                                                                   trans_c,
+                                                                   a_type,
+                                                                   b_type,
+                                                                   b1_type,
+                                                                   c_type,
+                                                                   ck_passthrough,
+                                                                   ck_passthrough,
+                                                                   ck_passthrough,
+                                                                   ck_passthrough};
+    }
+
+    operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
+    {
+        const auto& c_shape = inputs.back();
+        auto tuning_value   = v.get("tuning_value", 5);
+        auto batch_count    = get_batch_count(c_shape);
+        auto problem        = create_problem(inputs, v);
+
+        const auto include_header   = problem.GetIncludeHeader();
+        const auto solutions        = problem.GetSolutions(ctx.get_current_device().get_gfx_name());
+        const auto& solution        = solutions.at(tuning_value);
+        const auto template_str     = solution.template_str;
+        const auto blocks_per_batch = solution.grid_size;
+        const auto block_size       = solution.block_size;
+
+        hip_compile_options options;
+        options.additional_src_files = ck_headers();
+        auto grid_size = can_fold_batch(inputs) ? blocks_per_batch : batch_count * blocks_per_batch;
+        options.set_launch_params(v, grid_size * block_size, block_size);
+        options.inputs         = inputs;
+        options.output         = c_shape;
+        options.kernel_name    = v.get("kernel", "ck_gemm_softmax_gemm_kernel");
+        options.virtual_inputs = inputs;
+        if(can_fold_batch(inputs))
+        {
+            auto vinputs = inputs;
+            fold_batch_dims(vinputs[0]);
+            remove_batch_dims(vinputs[1]);
+            std::for_each(vinputs.begin() + 2, vinputs.end(), fold_batch_dims);
+            options.virtual_inputs = vinputs;
+        }
+
+        if(v.get("check", false) or enabled(MIGRAPHX_CK_DEBUG{}))
+            options.params += " -DMIGRAPHX_CK_CHECK=1";
+
+        // scale
+        assert(v.contains("scale"));
+        auto scale = v.at("scale").to<float>();
+        options.params += " -DSCALE=" + std::to_string(scale);
+
+        auto src = interpolate_string(ck_gemm_softmax_gemm_kernel,
+                                      {{"solution", template_str},
+                                       {"include", include_header},
+                                       {"params", enum_params(inputs.size(), "void * private_p")},
+                                       {"args", enum_params(inputs.size(), "private_p")},
+                                       {"blocks_per_batch", to_string(blocks_per_batch)},
+                                       {"preamble", v.get("preamble", std::string{})},
+                                       {"kernel", options.kernel_name}});
+
+        return compile_hip_code_object(src, options);
+    }
+
+    value create_settings(instruction_ref ins, const operation& op) const
+    {
+        auto v      = op.to_value();
+        v["kernel"] = "ck_gemm_softmax_gemm_kernel";
+        if(not ins->module_inputs().empty())
+        {
+            auto* pm      = ins->module_inputs().front();
+            v["preamble"] = generate_pointwise(*pm, "post_ck_gemm_softmax_gemm_function") +
+                            "\nMIGRAPHX_LIFT_CLASS(post_ck_gemm_softmax_gemm, "
+                            "post_ck_gemm_softmax_gemm_function);";
+            v["post"]   = "ck_function_adaptor<post_ck_gemm_softmax_gemm>";
+            v["kernel"] = "ck_gemm_softmax_gemm_" + generate_name_from_ops(*pm) + "_kernel";
+        }
+        return v;
+    }
+
+    compiler_replace
+    compile(context& ctx, instruction_ref ins, const operation& op, const value& solution) const
+    {
+        auto shapes = to_shapes(ins->inputs());
+        auto v      = create_settings(ins, op);
+        if(not solution.is_null())
+            v["tuning_value"] = solution;
+        return {compile_op(ctx, shapes, v),
+                [=](module& m, instruction_ref ins2, const operation& code_object) {
+                    if(enabled(MIGRAPHX_LOG_CK_GEMM{}))
+                    {
+                        std::vector<shape> gemm_shapes{
+                            shapes[0], shapes[1], shapes.back().with_type(shapes[0].type())};
+                        std::cout << "gpu::ck_gemm_softmax_gemm: "
+                                  << to_json_string(to_value(gemm_shapes)) << std::endl;
+                    }
+                    m.replace_instruction(ins2, code_object, ins2->inputs());
+                }};
+    }
+
+    optional<tuning_config>
+    get_tuning_config(context& ctx, instruction_ref ins, const operation& op, bool exhaustive) const
+    {
+        if(not exhaustive and not enabled(MIGRAPHX_TUNE_CK{}))
+            return nullopt;
+        tuning_config tc;
+        auto shapes    = to_shapes(ins->inputs());
+        auto problem   = create_problem(shapes, create_settings(ins, op));
+        auto solutions = problem.GetSolutions(ctx.get_current_device().get_gfx_name());
+        tc.solutions.resize(solutions.size());
+        std::iota(tc.solutions.begin(), tc.solutions.end(), 0);
+        std::vector<shape> gemm_shapes{shapes[0], shapes[1], shapes.back()};
+        tc.problem = to_value(gemm_shapes);
+        return tc;
+    }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/kernels/include/migraphx/kernels/ck.hpp

@@ -154,6 +154,17 @@ struct ck_add
     }
 };
 
+// In CK, the B matrix is ordered as N,K instead of K,N
+template <class Dims>
+constexpr auto ck_transposeb_dims(Dims dims)
+{
+    return unpack(dims, [](auto k, auto n) { return make_const_array(n, k); });
+}
+
+template <class Tensor>
+using ck_transposeb = decltype(make_shape(ck_transposeb_dims(get_shape_c<Tensor>{}.lens),
+                                          ck_transposeb_dims(get_shape_c<Tensor>{}.strides)));
+
 #ifdef MIGRAPHX_CK_CHECK
 #define MIGRAPHX_CK_STATIC_ASSERT static_assert
 #else

src/targets/gpu/kernels/include/migraphx/kernels/ck_gemm.hpp

@@ -33,17 +33,6 @@
 
 namespace migraphx {
 
-// In CK, the B matrix is ordered as N,K instead of K,N
-template <class Dims>
-constexpr auto ck_transposeb_dims(Dims dims)
-{
-    return unpack(dims, [](auto k, auto n) { return make_const_array(n, k); });
-}
-
-template <class Tensor>
-using ck_transposeb = decltype(make_shape(ck_transposeb_dims(get_shape_c<Tensor>{}.lens),
-                                          ck_transposeb_dims(get_shape_c<Tensor>{}.strides)));
-
 template <class G, class E, class A, class B, class... Ds>
 __device__ void ck_gemm_matrix(E e, A a, B b, Ds... ds)
 {

src/targets/gpu/kernels/include/migraphx/kernels/ck_gemm_softmax_gemm.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_KERNELS_CK_GEMM_SOFTMAX_GEMM_HPP
+#define MIGRAPHX_GUARD_KERNELS_CK_GEMM_SOFTMAX_GEMM_HPP
+
+#include <migraphx/kernels/index.hpp>
+#include <migraphx/kernels/algorithm.hpp>
+#include <migraphx/kernels/integral_constant.hpp>
+#include <migraphx/kernels/tensor_view.hpp>
+#include <migraphx/kernels/ck.hpp>
+#include <migraphx/kernels/gemm_batcher.hpp>
+
+namespace migraphx {
+
+template <class T>
+struct ck_gemm_softmax_gemm_settings
+{
+    T scale{};
+};
+
+template <class... Ts>
+constexpr ck_gemm_softmax_gemm_settings<Ts...> make_ck_gemm_softmax_gemm_settings(Ts... xs)
+{
+    return {xs...};
+}
+
+template <class G, class C, class A, class B, class B1, class Settings>
+__device__ void ck_gemm_softmax_gemm_matrix(C c, A a, B b, B1 b1, Settings s)
+{
+    constexpr auto desc = G::make_descriptor(to_ck_tensor<A>(),
+                                             to_ck_tensor<ck_transposeb<B>>(),
+                                             to_ck_tensor<ck_transposeb<B1>>(),
+                                             to_ck_tensor<C>());
+
+    static_assert(desc.IsValid(), "Invalid ck gemm.");
+
+    G::Run(desc,
+           s.scale,
+           to_ck_const_pointer(a.data()),
+           to_ck_const_pointer(b.data()),
+           to_ck_const_pointer(b1.data()),
+           to_ck_pointer(c.data()));
+}
+
+template <class G, index_int BlocksPerBatch, class... Ts, class Settings>
+__device__ void ck_gemm_softmax_gemm(Settings s, Ts... xs)
+{
+    gemm_batch_args(make_index(), _c<BlocksPerBatch>, xs...)(
+        [&](auto... ys) { ck_gemm_softmax_gemm_matrix<G>(ys..., s); });
+}
+
+} // namespace migraphx
+#endif

src/targets/gpu/kernels/include/migraphx/kernels/index.hpp

@@ -31,6 +31,14 @@
 #include <migraphx/kernels/debug.hpp>
 #include <migraphx/kernels/functional.hpp>
 
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wreserved-identifier"
+extern "C" __device__ size_t __ockl_get_enqueued_local_size(uint); // NOLINT
+extern "C" __device__ size_t __ockl_get_local_size(uint);          // NOLINT
+#pragma clang diagnostic pop
+#endif
+
 namespace migraphx {
 
 #if defined(MIGRAPHX_NGLOBAL) && defined(MIGRAPHX_NLOCAL)
@@ -45,43 +53,37 @@ inline __device__ __attribute__((const)) index_int compute_global_size()
     // This actualy works even when global is not divisible by local size.
     // This doesnt actually do a multiplicatiosn. Instead it calls a device
     // function to get the global size, which is why it works.
-    return blockDim.x * gridDim.x;  // NOLINT
+    return blockDim.x * gridDim.x; // NOLINT
 #endif
 }
 
-// We cant just use blockDim.x to get the local size since its broken on hip
-// when global is not divisible by local size. In this case, we calulate the
-// size for the last group.
+#ifdef MIGRAPHX_NGROUP
+// If global is divisible by local then local can be a const
+#if(MIGRAPHX_NGLOBAL % MIGRAPHX_NLOCAL == 0) || (MIGRAPHX_NGROUP == 1)
+#define MIGRAPHX_HAS_CONST_LOCAL 1
+#endif
+#endif
+
 inline __device__ __attribute__((const)) index_int compute_local_size()
 {
-#ifdef MIGRAPHX_NLOCAL
-    const auto nlocal = MIGRAPHX_NLOCAL;
-#else
-    const auto nlocal = blockDim.x; // NOLINT
-#endif
-#ifdef MIGRAPHX_NGROUP
-    const auto ngroup = MIGRAPHX_NGROUP;
+#ifdef MIGRAPHX_HAS_CONST_LOCAL
+    return MIGRAPHX_NLOCAL;
 #else
-    const auto ngroup = gridDim.x;  // NOLINT
+    // Returns block size. For the non-uniform block it returns the size of the non-uniform block.
+    return __ockl_get_local_size(0); // NOLINT
 #endif
-    const auto group_id = blockIdx.x; // NOLINT
-    const auto nglobal  = compute_global_size();
-    if(group_id == ngroup - 1)
-    {
-        return 1 + (nglobal - 1) % nlocal;
-    }
-    else
-    {
-        return nlocal; // NOLINT
-    }
 }
 
-#ifdef MIGRAPHX_NGROUP
-// If global is divisible by local then local can be a const
-#if(MIGRAPHX_NGLOBAL % MIGRAPHX_NLOCAL == 0) || (MIGRAPHX_NGROUP == 1)
-#define MIGRAPHX_HAS_CONST_LOCAL 1
-#endif
+inline __device__ __attribute__((const)) index_int compute_max_local_size()
+{
+#ifdef MIGRAPHX_LOCAL
+    return MIGRAPHX_NLOCAL;
+#else
+    // Returns the block size. When workgrop has non-uniform block, this returns size of the uniform
+    // block.
+    return __ockl_get_enqueued_local_size(0); // NOLINT
 #endif
+}
 
 struct index
 {
@@ -126,8 +128,8 @@ struct index
 #else
     __device__ index_int max_nlocal() const
     {
-        MIGRAPHX_ASSERT(blockDim.x > 0);
-        return blockDim.x;
+        MIGRAPHX_ASSERT(compute_max_local_size() > 0);
+        return compute_max_local_size();
     }
 #endif
 
@@ -249,7 +251,8 @@ struct index
 #endif
 inline __device__ __attribute__((const)) index make_index()
 {
-    return index{blockIdx.x * blockDim.x + threadIdx.x, threadIdx.x, blockIdx.x}; // NOLINT
+    return index{
+        blockIdx.x * compute_max_local_size() + threadIdx.x, threadIdx.x, blockIdx.x}; // NOLINT
 }
 
 } // namespace migraphx

src/targets/gpu/mlir.cpp

@@ -320,7 +320,10 @@ struct mlir_program
 
     MlirType make_tensor(const shape& s) const
     {
-        assert(s.standard());
+        if(not s.standard())
+            MIGRAPHX_THROW("MLIR expects all tensors to be in standard shape");
+        if(s.dynamic())
+            MIGRAPHX_THROW("MLIR does not support dynamic shapes");
         std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
         return mlirRankedTensorTypeGet(
             lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());

src/targets/gpu/prefuse_ops.cpp

@@ -23,16 +23,17 @@
  */
 #include <migraphx/permutation.hpp>
 #include <migraphx/gpu/prefuse_ops.hpp>
+#include <migraphx/gpu/gemm_softmax_gemm.hpp>
 #include <migraphx/match/layernorm.hpp>
-#include <migraphx/check_shapes.hpp>
-#include <migraphx/make_op.hpp>
 #include <migraphx/register_op.hpp>
 #include <migraphx/pass_manager.hpp>
 #include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/gpu/ck.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
+
 namespace {
 
 template <class Derived, std::size_t N>
@@ -120,6 +121,60 @@ struct find_add_layernorm
         m.replace_instruction(ins, add_layernorm{op.epsilon}, add_ins->inputs());
     }
 };
+
+struct pre_gemm_softmax_gemm : gemm_softmax_gemm
+{
+    std::string name() const { return "gpu::pre_gemm_softmax_gemm"; }
+};
+MIGRAPHX_REGISTER_OP(pre_gemm_softmax_gemm);
+
+MIGRAPHX_PRED_MATCHER(is_ck_gemm, instruction_ref ins)
+{
+    if(ins->name() != "dot")
+        return false;
+    if(not pre_gemm_softmax_gemm::is_ck_supported_type(ins->get_shape().type()))
+        return false;
+    return true;
+}
+
+struct find_gemm_softmax_gemm
+{
+    auto matcher() const
+    {
+        auto gemm1 =
+            match::skip(match::name("contiguous"))(match::name("dot")(is_ck_gemm().bind("gemm1")));
+        auto mul = match::name("mul")(
+            match::nargs(2), match::either_arg(0, 1)(match::is_constant().bind("scale"), gemm1));
+        auto softmax = match::name("softmax")(match::arg(0)(mul)).bind("softmax");
+
+        return match::name("dot")(is_ck_gemm().bind("gemm2"))(match::arg(0)(softmax));
+    }
+
+    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
+    {
+        auto ins       = r.result;
+        auto gemm2_ins = r.instructions["gemm2"];
+        auto gemm1_ins = r.instructions["gemm1"];
+        auto scale_lit = r.instructions["scale"];
+
+        float scale = 1.0;
+        scale_lit->eval().visit([&](const auto s) {
+            // CK only supports single-valued scale
+            if(std::all_of(
+                   s.begin() + 1, s.end(), [&](auto v) { return float_equal(v, s.front()); }))
+                scale = s.front();
+            else
+                return;
+        });
+
+        auto inputs = gemm1_ins->inputs();            // A, B
+        inputs.push_back(gemm2_ins->inputs().back()); // B1
+
+        mpm.get_module().replace_instruction(
+            ins, pre_gemm_softmax_gemm{gemm2_ins->get_operator(), scale}, inputs);
+    }
+};
+
 } // namespace
 
 void prefuse_ops::apply(module_pass_manager& mpm) const
@@ -127,6 +182,8 @@ void prefuse_ops::apply(module_pass_manager& mpm) const
     match::find_matches(mpm.get_module(), find_layernorm{});
     mpm.run_pass(dead_code_elimination{});
     match::find_matches(mpm.get_module(), find_add_layernorm{});
+    if(enabled(MIGRAPHX_ENABLE_CK{}))
+        match::find_matches(mpm, find_gemm_softmax_gemm{});
 }
 
 } // namespace gpu

src/targets/gpu/time_op.cpp

@@ -41,8 +41,7 @@ std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsig
 }
 
 using milliseconds = std::chrono::duration<double, std::milli>;
-std::pair<double, double>
-time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
+double time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
 {
 
     // TODO: Use std::ref
@@ -51,21 +50,19 @@ time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
     auto output           = op.compute_shape(inputs);
     op.finalize(ctx, output, inputs);
     auto args = generate_arguments(inputs);
-    auto run  = [&] {
-        op.compute(ctx, output, args);
-        ctx.finish();
-    };
-    gctx.enable_perf_measurement();
+    auto start = context::create_event_for_timing();
+    auto stop  = context::create_event_for_timing();
+    auto run   = [&] { op.compute(ctx, output, args); };
     run();
-    double host_time   = 0.0;
-    double device_time = 0.0;
+    gctx.get_stream().record(start.get());
     for(auto i : range(n))
     {
         (void)i;
-        host_time += time<milliseconds>(run);
-        device_time += gctx.get_elapsed_ms();
+        run();
     }
-    return std::make_pair(host_time / n, device_time / n);
+    gctx.get_stream().record(stop.get());
+    gctx.finish();
+    return context::get_elapsed_ms(start.get(), stop.get()) / n;
 }
 
 } // namespace gpu

src/tf/parse_reshape.cpp

@@ -45,8 +45,7 @@ struct parse_reshape : op_parser<parse_reshape>
         auto s = args[1]->eval();
         std::vector<int64_t> dims;
         s.visit([&](auto v) { copy(v, std::back_inserter(dims)); });
-        return info.add_instruction(make_op("reshape", {{"dims", dims}}),
-                                    info.make_contiguous(args[0]));
+        return info.add_instruction(make_op("reshape", {{"dims", dims}}), args[0]);
     }
 };
 

test/eliminate_allocation_test.cpp

@@ -55,7 +55,7 @@ struct allocate
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 

test/eliminate_concat_test.cpp

@@ -60,7 +60,7 @@ struct concat
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 
@@ -104,7 +104,7 @@ struct allocate
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 

test/gpu/fuse_mlir.cpp

@@ -34,7 +34,8 @@
 
 void run_pass(migraphx::program& p)
 {
-    migraphx::run_passes(p, {migraphx::gpu::fuse_mlir{}, migraphx::dead_code_elimination{}});
+    migraphx::run_passes(
+        p, {migraphx::gpu::fuse_mlir{.enable_extra = true}, migraphx::dead_code_elimination{}});
 }
 
 template <class F>
@@ -151,7 +152,6 @@ TEST_CASE(int_quant_dot_tanh_fails)
 
 int main(int argc, const char* argv[])
 {
-    if(migraphx::gpu::mlir_enabled())
-        test::run(argc, argv);
+    test::run(argc, argv);
     return 0;
 }