Skip to content

GitLab

Commit 9032352f authored by Jakub Piasecki's avatar Jakub Piasecki
Browse files

resolved conflicts

parents d5c5d2a3 64d5c4d6
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 805 additions and 410 deletions
example/ck_tile/03_gemm/instances/gemm_universal_mem_instance_common.hpp
View file @ 9032352f
...@@ -14,7 +14,7 @@ float gemm_(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s) ...@@ -14,7 +14,7 @@ float gemm_(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
ck_tile::sequence<Traits_::M_Tile, Traits_::N_Tile, Traits_::K_Tile>, ck_tile::sequence<Traits_::M_Tile, Traits_::N_Tile, Traits_::K_Tile>,
ck_tile::sequence<Traits_::M_Warp, Traits_::N_Warp, Traits_::K_Warp>, ck_tile::sequence<Traits_::M_Warp, Traits_::N_Warp, Traits_::K_Warp>,
ck_tile::sequence<Traits_::M_Warp_Tile, Traits_::N_Warp_Tile, Traits_::K_Warp_Tile>>; ck_tile::sequence<Traits_::M_Warp_Tile, Traits_::N_Warp_Tile, Traits_::K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>; using TilePartitioner = ck_tile::GemmTile2DPartitioner<GemmShape>;
using GemmEpilogue = using GemmEpilogue =
ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename Traits_::AccDataType, ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename Traits_::AccDataType,
......
example/ck_tile/03_gemm/run_gemm_example.inc
View file @ 9032352f
...@@ -25,6 +25,33 @@ auto create_args(int argc, char* argv[]) ...@@ -25,6 +25,33 @@ auto create_args(int argc, char* argv[])
return std::make_tuple(result, arg_parser); return std::make_tuple(result, arg_parser);
} }
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
{
return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
ck_tile::tensor_layout::gemm::RowMajor>>{};
}
auto calculate_rtol_atol(const ck_tile::index_t K,
const ck_tile::index_t kbatch,
const float max_accumulated_value)
{
using ComputeType =
std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>;
// Calculate thresholds
const auto rtol = ck_tile::get_relative_threshold<ComputeType, CDataType, AccDataType>(
ck_tile::integer_divide_ceil(K, kbatch));
const auto atol = ck_tile::get_absolute_threshold<ComputeType, CDataType, AccDataType>(
max_accumulated_value / kbatch, ck_tile::integer_divide_ceil(K, kbatch));
// Calculate error due to split_k accumulation
const auto rtol_split_k =
ck_tile::get_relative_threshold<CDataType, CDataType, CDataType>(kbatch);
const auto atol_split_k = ck_tile::get_absolute_threshold<CDataType, CDataType, CDataType>(
max_accumulated_value, kbatch);
// Use higher threshold
return ck_tile::make_tuple(std::max(rtol, rtol_split_k), std::max(atol, atol_split_k));
}
template <typename ALayout, typename BLayout, typename CLayout> template <typename ALayout, typename BLayout, typename CLayout>
float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf, float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::DeviceMem& b_k_n_dev_buf, ck_tile::DeviceMem& b_k_n_dev_buf,
...@@ -97,48 +124,16 @@ int run_gemm_example_with_layouts(int argc, ...@@ -97,48 +124,16 @@ int run_gemm_example_with_layouts(int argc,
int n_warmup = arg_parser.get_int("warmup"); int n_warmup = arg_parser.get_int("warmup");
int n_repeat = arg_parser.get_int("repeat"); int n_repeat = arg_parser.get_int("repeat");
using namespace ck_tile::literals; stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
auto f_host_tensor_descriptor = stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>) ck_tile::HostTensor<ADataType> a_m_k(
{ ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
return ck_tile::HostTensorDescriptor({row, col}, {stride, 1_uz}); ck_tile::HostTensor<BDataType> b_k_n(
} ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
else
{
return ck_tile::HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride = [](std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if(stride == 0)
{
// give a chance if stride is zero, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return col;
}
else
{
return row;
}
}
else
return stride;
};
stride_A = f_get_default_stride(M, K, stride_A, a_layout);
stride_B = f_get_default_stride(K, N, stride_B, b_layout);
stride_C = f_get_default_stride(M, N, stride_C, CLayout{});
ck_tile::HostTensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, stride_A, a_layout));
ck_tile::HostTensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, stride_B, b_layout));
ck_tile::HostTensor<CDataType> c_m_n_dev_result( ck_tile::HostTensor<CDataType> c_m_n_dev_result(
f_host_tensor_descriptor(M, N, stride_C, CLayout{})); ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
// TODO: add different init types // TODO: add different init types
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k); ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
...@@ -172,20 +167,29 @@ int run_gemm_example_with_layouts(int argc, ...@@ -172,20 +167,29 @@ int run_gemm_example_with_layouts(int argc,
if(arg_parser.get_int("v") == 1) if(arg_parser.get_int("v") == 1)
{ {
ck_tile::HostTensor<CDataType> c_m_n_host_ref( ck_tile::HostTensor<CDataType> c_m_n_host_ref(
f_host_tensor_descriptor(M, N, stride_C, CLayout{})); ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
c_m_n_host_ref.SetZero(); c_m_n_host_ref.SetZero();
ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>( ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
a_m_k, b_k_n, c_m_n_host_ref); a_m_k, b_k_n, c_m_n_host_ref);
const float max_accumulated_value =
pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_host_ref); *std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_host_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
<< std::endl;
std::cout << "The CPU veification result is:" << (pass ? "correct" : "fail") << std::endl; std::cout << "The CPU veification result is:" << (pass ? "correct" : "fail") << std::endl;
} }
else if(arg_parser.get_int("v") == 2) else if(arg_parser.get_int("v") == 2)
{ {
ck_tile::HostTensor<CDataType> c_m_n_gpu_ref( ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(
f_host_tensor_descriptor(M, N, stride_C, CLayout{})); ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes()); ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes());
c_m_n_gpu_ref.SetZero(); c_m_n_gpu_ref.SetZero();
c_m_n_gpu_buf_ref.SetZero(); c_m_n_gpu_buf_ref.SetZero();
...@@ -225,8 +229,18 @@ int run_gemm_example_with_layouts(int argc, ...@@ -225,8 +229,18 @@ int run_gemm_example_with_layouts(int argc,
ck_tile::hip_check_error(hipFree(d_C)); ck_tile::hip_check_error(hipFree(d_C));
c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data()); c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_gpu_ref); const float max_accumulated_value =
*std::max_element(c_m_n_gpu_ref.mData.begin(), c_m_n_gpu_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_gpu_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
<< std::endl;
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl; std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
} }
......
example/ck_tile/10_rmsnorm2d/CMakeLists.txt
View file @ 9032352f
set(RMSNORM2D_FWD_KNOWN_APIS "fwd;bwd")
set(RMSNORM2D_FWD_ENABLE_APIS "fwd" CACHE STRING
"semicolon-separated list of APIs to generate (${RMSNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
if(RMSNORM2D_FWD_ENABLE_APIS STREQUAL "all")
set(RMSNORM2D_FWD_ENABLE_APIS ${RMSNORM2D_FWD_KNOWN_APIS})
endif()
# generate a list of kernels, but not actually emit files at config sta
execute_process(
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
RESULT_VARIABLE ret
)
if(ret AND NOT ret EQUAL 0)
message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
endif()
file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/rmsnorm2d_fwd_blobs.txt RMSNORM2D_FWD_GEN_BLOBS)
add_custom_command(
OUTPUT ${RMSNORM2D_FWD_GEN_BLOBS}
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
)
set(TILE_RMSNORM2D_FWD "tile_rmsnorm2d_fwd") set(TILE_RMSNORM2D_FWD "tile_rmsnorm2d_fwd")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message("adding ${TILE_RMSNORM2D_FWD}") message("adding ${TILE_RMSNORM2D_FWD}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${TILE_RMSNORM2D_FWD} EXCLUDE_FROM_ALL rmsnorm2d_fwd.cpp) add_executable(${TILE_RMSNORM2D_FWD} EXCLUDE_FROM_ALL rmsnorm2d_fwd.cpp)
target_include_directories(${TILE_RMSNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR}) target_include_directories(${TILE_RMSNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${TILE_RMSNORM2D_FWD} PRIVATE ${INSTANCE_SRCS}) target_sources(${TILE_RMSNORM2D_FWD} PRIVATE ${RMSNORM2D_FWD_GEN_BLOBS})
set(TILE_RMSNORM2D_FWD_COMPILE_OPTIONS) set(TILE_RMSNORM2D_FWD_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations # NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND TILE_RMSNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal) list(APPEND TILE_RMSNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal --offload-compress)
target_compile_options(${TILE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_FWD_COMPILE_OPTIONS}) target_compile_options(${TILE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_FWD_COMPILE_OPTIONS})
......
example/ck_tile/10_rmsnorm2d/example_rmsnorm2d_fwd.cpp
View file @ 9032352f
#include "ck_tile/host.hpp" #include "ck_tile/host.hpp"
#include "ck_tile/core.hpp" #include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp" #include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/rmsnorm2d.hpp" #include "ck_tile/ops/rmsnorm2d.hpp"
#include <cstring> #include <cstring>
...@@ -36,10 +37,12 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -36,10 +37,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
assert(stride >= n); assert(stride >= n);
using XDataType = DataType; using XDataType = DataType;
using YDataType = DataType; using YDataType = DataType;
using GammaDataType = DataType; using GammaDataType = DataType;
using InvRmsDataType = ck_tile::null_type; using InvRmsDataType = ck_tile::null_type;
using SmoothScaleDataType = ck_tile::null_type;
using YScaleDataType = ck_tile::null_type;
using ComputeDataType = float; using ComputeDataType = float;
...@@ -68,30 +71,49 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -68,30 +71,49 @@ bool run(const ck_tile::ArgParser& arg_parser)
using BlockTile = ck_tile::sequence<2, 128>; using BlockTile = ck_tile::sequence<2, 128>;
using WarpTile = ck_tile::sequence<1, 64>; using WarpTile = ck_tile::sequence<1, 64>;
using Vector = ck_tile::sequence<1, 1>; using Vector = ck_tile::sequence<1, 1>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using PipelineTraits =
ck_tile::Rmsnorm2dFwdTraits<true, // kPadN
false, // kSaveInvRms
kTwoPass,
ck_tile::Rmsnorm2dFusedAddEnum::NO_ADD, // fuse add
ck_tile::Rmsnorm2dFusedQuantEnum::NO_SWEEP>; // fuse quant
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using Problem = ck_tile::Rmsnorm2dFwdPipelineProblem<XDataType, using Problem = ck_tile::Rmsnorm2dFwdPipelineProblem<XDataType,
GammaDataType, GammaDataType,
ComputeDataType, ComputeDataType,
YDataType, YDataType,
InvRmsDataType, InvRmsDataType,
SmoothScaleDataType,
YScaleDataType,
Shape, Shape,
true, // kPadN PipelineTraits>;
false, // kSaveInvRms
kTwoPass>;
using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<Problem>; using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<Problem>;
using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<Problem>; using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<Problem>;
using Pipeline = std::conditional_t<kTwoPass, TwoPassPipeline, OnePassPipeline>; using Pipeline = std::conditional_t<kTwoPass, TwoPassPipeline, OnePassPipeline>;
using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline>;
using Default2DEpilogueProblem = ck_tile::
Default2DEpilogueProblem<ComputeDataType, YDataType, false, PipelineTraits::kPadN, false>;
using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline, Default2DEpilogue>;
ck_tile::Rmsnorm2dFwdHostArgs args{x_buf.GetDeviceBuffer(), ck_tile::Rmsnorm2dFwdHostArgs args{x_buf.GetDeviceBuffer(),
nullptr,
nullptr,
gamma_buf.GetDeviceBuffer(), gamma_buf.GetDeviceBuffer(),
y_buf.GetDeviceBuffer(), y_buf.GetDeviceBuffer(),
nullptr, nullptr,
nullptr,
nullptr,
epsilon, epsilon,
m, m,
n, n,
stride,
stride,
stride,
stride}; stride};
auto kargs = Kernel::MakeKargs(args); auto kargs = Kernel::MakeKargs(args);
......
example/ck_tile/10_rmsnorm2d/generate.py 0 → 100644
View file @ 9032352f
This diff is collapsed.
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_api.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "rmsnorm2d_fwd.hpp"
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 kSaveInvRms_,
bool kTwoPass_>
using trait_ = rmsnorm2d_fwd_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kSaveInvRms_,
kTwoPass_>;
template <typename data_type>
float rmsnorm2d_fwd_b16_(rmsnorm2d_fwd_traits /*t*/,
rmsnorm2d_fwd_args a,
const ck_tile::stream_config& s)
{
float r = -1;
// clang-format off
// rm rn tm tn vn pd rms 2p
if(a.n <= 64) {
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 128) {
if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 256) {
if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 512) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 8, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 768) {
if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 6, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1,12, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 1024) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 2, 128, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 2, 128, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 2, 128, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 1536) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 4, 64, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 2, 128, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 256, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 6, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 2048) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 1, 256, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 8, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 3072) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 128, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 6, 1, 256, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 1024, 1, true, false, false>>(s, a);
}
else if(a.n <= 4096) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 1024, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 1024, 1, true, false, false>>(s, a);
}
else if(a.n > 4096) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 8, true, false, true>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 4, true, false, true>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 1024, 2, true, false, true>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 1024, 1, true, false, true>>(s, a);
}
return r;
// clang-format on
}
float rmsnorm2d_fwd(rmsnorm2d_fwd_traits t, rmsnorm2d_fwd_args a, const ck_tile::stream_config& s)
{
if(t.data_type.compare("fp16") == 0)
{
return rmsnorm2d_fwd_b16_<ck_tile::fp16_t>(t, a, s);
}
else if(t.data_type.compare("bf16") == 0)
{
return rmsnorm2d_fwd_b16_<ck_tile::bf16_t>(t, a, s);
}
else
throw std::runtime_error("Without supported instances!");
}
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n1024_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
#if 0
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n1536_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n2048_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n256_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n3072_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n4096_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n4096_tp_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, true>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n512_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n64_n128_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 1, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n768_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 12, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n1024_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
#if 0
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n1536_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n2048_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n256_instance.cpp deleted 100644 → 0
View file @ d5c5d2a3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment