ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 10:37:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit 'ec23be0b9d45ff9ca4135090bcd0269184c953a7' into develop

Browse Source
This commit is contained in:
assistant-librarian[bot]
2026-01-03 06:16:07 +00:00
parent 0b05cd0351
commit d8734393c7
6 changed files with 22 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
example/ck_tile/13_moe_sorting/moe_sorting.cpp
View File

@@ -334,13 +334,13 @@ bool test_moe_sorting(ck_tile::ArgParser args)
if(moe_buf_bytes > 0)
{
#if MOE_SORTING_FMOE_2D_BUF
printf("moe_buf:%lu(%d,%d), ",
printf("moe_buf:%" PRIu64 "(%d,%d), ",
static_cast<uint64_t>(moe_buf_bytes),
moe_buf_interm_dim,
moe_buf_elem_bytes);
#else
printf("moe_buf:%lu, ", static_cast<uint64_t>(moe_buf_bytes));
printf("moe_buf:%" PRIu64 ", ", static_cast<uint64_t>(moe_buf_bytes));
#endif
}

7
include/ck_tile/host/fill.hpp
View File

@@ -55,9 +55,10 @@ struct FillUniformDistribution
const auto total_bytes = total * sizeof(T_iter);
// max 80 threads; at least 2MB per thread
const size_t available_cpu_cores = get_available_cpu_cores();
const size_t num_thread =
min(80UL, available_cpu_cores, integer_divide_ceil(total_bytes, 0x200000UL));
const size_t available_cpu_cores = get_available_cpu_cores();
constexpr uint64_t MAX_THREAD_COUNT = 80;
const size_t num_thread = min(
MAX_THREAD_COUNT, available_cpu_cores, integer_divide_ceil(total_bytes, 0x200000UL));
constexpr size_t BLOCK_BYTES = 64;
constexpr size_t BLOCK_SIZE = BLOCK_BYTES / sizeof(T_iter);
const size_t num_blocks = integer_divide_ceil(total_bytes, BLOCK_BYTES);

3
include/ck_tile/host/reference/reference_grouped_conv_bwd_data.hpp
View File

@@ -3,6 +3,7 @@
#pragma once
#include <cinttypes>
#include <cstdlib>
#include <thread>
@@ -28,7 +29,7 @@ CK_TILE_HOST void reference_grouped_conv_bwd_data(HostTensor<InDataType>& input,
output.get_num_of_dimension() == NDimSpatial + 3))
{
printf("%lu %lu %lu",
printf("%" PRIu64 " %" PRIu64 " %" PRIu64,
input.get_num_of_dimension(),
weight.get_num_of_dimension(),
output.get_num_of_dimension());

16
include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp
View File

@@ -246,9 +246,11 @@ struct UniversalGemmBasePolicy
}
else // A is in RowMajor
{
constexpr auto DataTypeSize = sizeof(ADataType);
constexpr auto DataTypeSize = sizeof(ADataType);
constexpr uint64_t MinLdsLayer = 1ULL;
constexpr auto MLdsLayer =
max(1UL, get_n_lds_banks() * get_n_words_per_128b() / KPerBlock / DataTypeSize);
max(MinLdsLayer,
get_n_lds_banks() * get_n_words_per_128b() / KPerBlock / DataTypeSize);
constexpr index_t NBanks = get_n_lds_banks();
static_assert(NBanks == 32 || NBanks == 64, "Unexpected LDS bank count");
@@ -442,11 +444,13 @@ struct UniversalGemmBasePolicy
}
else // B is Column Major
{
constexpr index_t KPack = GetSmemPackB<Problem>();
constexpr auto BK0 = number<KPerBlock / KPack>{};
constexpr auto DataTypeSize = sizeof(BDataType);
constexpr index_t KPack = GetSmemPackB<Problem>();
constexpr auto BK0 = number<KPerBlock / KPack>{};
constexpr auto DataTypeSize = sizeof(BDataType);
constexpr uint64_t MinLdsLayer = 1ULL;
constexpr auto NLdsLayer =
max(1UL, get_n_lds_banks() * get_n_words_per_128b() / KPerBlock / DataTypeSize);
max(MinLdsLayer,
get_n_lds_banks() * get_n_words_per_128b() / KPerBlock / DataTypeSize);
constexpr index_t NBanks = get_n_lds_banks();
static_assert(NBanks == 32 || NBanks == 64, "Unexpected LDS bank count");

4
test/ck_tile/moe_sorting/test_moe_sorting_util.hpp
View File

@@ -236,13 +236,13 @@ class TestCkTileMoeSorting : public ::testing::Test
if(moe_buf_bytes > 0)
{
#if MOE_SORTING_FMOE_2D_BUF
printf("moe_buf:%lu(%d,%d), ",
printf("moe_buf:%" PRIu64 "(%d,%d), ",
static_cast<uint64_t>(moe_buf_bytes),
moe_buf_interm_dim,
moe_buf_elem_bytes);
#else
printf("moe_buf:%lu, ", static_cast<uint64_t>(moe_buf_bytes));
printf("moe_buf:%" PRIu64 ", ", static_cast<uint64_t>(moe_buf_bytes));
#endif
}

2
test/ck_tile/utility/test_fill.cpp
View File

@@ -26,6 +26,7 @@ using TestTypes = ::testing::Types<float, fp16_t, fp8_t, pk_fp4_t>;
TYPED_TEST_SUITE(FillUniformDistributionTest, TestTypes);
// Test that multiple runs with the same seed produce identical results
#ifndef _WIN32
TYPED_TEST(FillUniformDistributionTest, ConsistencyWithSameSeed)
{
using T = TypeParam;
@@ -53,6 +54,7 @@ TYPED_TEST(FillUniformDistributionTest, ConsistencyWithSameSeed)
<< "First and second fill should be identical";
}
}
#endif
// Test consistency across different data sizes (which affects threading)
TYPED_TEST(FillUniformDistributionTest, ConsistencyAcrossSizes)