Tests for CK Tile Batched Transpose and Smoothquant (#2453)

* Create tests for ck tile batched transpose using example * Create ck tile tests for smoothquant using examples * fix precision input strings and convert batched transpose to regression tests * Code cleanup and fix asserts * add missing licenses * update copyright and licensing in files * Update smoothquant tests to use example's smoothquant.cpp * Add custom target for batched transpose tests * Add missing new lines at end of files for CMakelists * fix typo in batched transpose CMakeList target_compile_options --------- Co-authored-by: root <root@ctr-ubbsmc16.amd.com>
2026-04-19 22:39:03 +00:00 · 2025-07-17 09:53:34 -06:00
parent 7fc000d7b3
commit c08986b026
36 changed files with 1391 additions and 0 deletions
--- a/test/ck_tile/smoothquant/instances/smoothquant_instance_common.hpp
+++ b/test/ck_tile/smoothquant/instances/smoothquant_instance_common.hpp
@@ -0,0 +1,61 @@
+// Copyright © Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier:  MIT
+
+#include <ck_tile/core.hpp>
+#include "smoothquant.hpp"
+#include <iostream>
+
+#pragma once
+
+using S = ck_tile::stream_config;
+using A = smoothquant_args;
+
+template <typename DataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kTwoPass_>
+using trait_ = smoothquant_traits_<DataType_,
+                                   Repeat_M_,
+                                   Repeat_N_,
+                                   ThreadPerBlock_M_,
+                                   ThreadPerBlock_N_,
+                                   Vector_N_,
+                                   kPadN_,
+                                   kTwoPass_>;
+
+template <typename Traits_>
+float smoothquant_(const S& s, A a)
+{
+    using DataType = typename Traits_::DataType;
+
+    using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
+        typename SmoothquantTypeConfig<DataType>::XDataType,
+        typename SmoothquantTypeConfig<DataType>::SmoothScaleDataType,
+        typename SmoothquantTypeConfig<DataType>::ComputeDataType,
+        typename SmoothquantTypeConfig<DataType>::YScaleDataType,
+        typename SmoothquantTypeConfig<DataType>::QYDataType,
+        typename Traits_::Shape,
+        Traits_::kPadN,
+        Traits_::kTwoPass>;
+
+    using OnePassPipeline = ck_tile::SmoothquantPipelineOnePass<PipelineProblem>;
+    using TwoPassPipeline = ck_tile::SmoothquantPipelineTwoPass<PipelineProblem>;
+    using Pipeline        = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
+
+    using Kernel = ck_tile::Smoothquant<Pipeline>;
+
+    const dim3 grids                       = Kernel::GridSize(a);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+
+    auto kargs = Kernel::MakeKargs(a);
+    if(s.log_level_ > 0)
+        std::cout << ", " << Kernel::GetName() << std::flush;
+
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+}