Skip to content

GitLab

Commit 6e3c786e authored by Jing Zhang's avatar Jing Zhang
Browse files

merge develop

parents 1bb510cb 261f1759
Hide whitespace changes
Inline Side-by-side
Showing with 291 additions and 94 deletions
example/ck_tile/CMakeLists.txt
View file @ 6e3c786e
......@@ -6,3 +6,14 @@ add_subdirectory(01_fmha)
add_subdirectory(02_layernorm2d)
add_subdirectory(03_gemm)
add_subdirectory(04_img2col)
add_subdirectory(05_reduce)
add_subdirectory(06_permute)
add_subdirectory(09_topk_softmax)
add_subdirectory(10_rmsnorm2d)
add_subdirectory(11_add_rmsnorm2d_rdquant)
add_subdirectory(12_smoothquant)
add_subdirectory(13_moe_sorting)
add_subdirectory(14_moe_smoothquant)
add_subdirectory(15_fused_moe)
add_subdirectory(16_batched_gemm)
add_subdirectory(17_grouped_gemm)
include/ck/README.md 0 → 100644
View file @ 6e3c786e
[Back to the main page](../../README.md)
# Composable Kernel supported operations
## Supported device operations
* [Average pooling]()
* [Batched contraction]()
* [Batched gemm]()
* [Batchnorm]()
* [CGEMM]()
* [Contraction]()
* [Convolution]()
* [Image to Column and Column to Image]()
* [Elementwise]()
* [GEMM]()
* [Max pooling]()
* [Reduce]()
* [Normalization]()
* [Permute]()
* [Put]()
* [Softmax]()
include/ck/ck.hpp
View file @ 6e3c786e
......@@ -63,13 +63,15 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
#define __gfx101__
#endif
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__)
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || \
defined(__gfx10_3_generic__)
#define __gfx103__
#endif
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || \
defined(__gfx1103__) || defined(__gfx11_generic__)
#define __gfx11__
#endif
#if defined(__gfx1200__) || defined(__gfx1201__)
#if defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx12_generic__)
#define __gfx12__
#endif
......
include/ck/host_utility/flush_cache.hpp
View file @ 6e3c786e
......@@ -237,7 +237,7 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
Args... args)
{
#if CK_TIME_KERNEL
#define MEDIAN 1
#define MEDIAN 0
if(stream_config.time_kernel_)
{
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
......@@ -275,6 +275,14 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
#else
float total_time = 0;
#endif
hipEvent_t start, stop;
hip_check_error(hipEventCreate(&start));
hip_check_error(hipEventCreate(&stop));
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(start, stream_config.stream_id_));
for(int i = 0; i < nrepeat; ++i)
{
if constexpr(!TimePreprocess)
......@@ -282,13 +290,13 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
preprocess();
}
hipEvent_t start, stop;
// hipEvent_t start, stop;
hip_check_error(hipEventCreate(&start));
hip_check_error(hipEventCreate(&stop));
// hip_check_error(hipEventCreate(&start));
// hip_check_error(hipEventCreate(&stop));
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(start, stream_config.stream_id_));
// hip_check_error(hipDeviceSynchronize());
// hip_check_error(hipEventRecord(start, stream_config.stream_id_));
// calculate preprocess time
if constexpr(TimePreprocess)
{
......@@ -299,25 +307,34 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
hip_check_error(hipGetLastError());
// end real kernel
hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
hip_check_error(hipEventSynchronize(stop));
float cur_time = 0;
hip_check_error(hipEventElapsedTime(&cur_time, start, stop));
#if MEDIAN
times.insert(cur_time);
#else
total_time += cur_time;
#endif
// hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
// hip_check_error(hipEventSynchronize(stop));
// float cur_time = 0;
// hip_check_error(hipEventElapsedTime(&cur_time, start, stop));
// #if MEDIAN
// times.insert(cur_time);
// #else
// total_time += cur_time;
// #endif
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
std::cout << "i: " << i << " cur_time: " << cur_time << std::endl;
// std::cout << "i: " << i << " cur_time: " << cur_time << std::endl;
printf("gemm_args.p_a_grid: %p, gemm_args.p_b_grid:%p\n",
static_cast<const void*>(gemm_args.p_a_grid),
static_cast<const void*>(gemm_args.p_b_grid));
}
}
hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
hip_check_error(hipEventSynchronize(stop));
float cur_time = 0;
hip_check_error(hipEventElapsedTime(&cur_time, start, stop));
#if MEDIAN
times.insert(cur_time);
#else
total_time += cur_time;
#endif
#if MEDIAN
auto mid = times.begin();
......@@ -333,7 +350,11 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
return (*mid + *mid_next) / 2;
}
#else
return total_time / nrepeat;
// return total_time / nrepeat;
hipDeviceProp_t deviceProps;
hip_check_error(hipGetDeviceProperties(&deviceProps, 0));
float preprocess_offset = deviceProps.multiProcessorCount == 80 ? 0.005 : 0.01;
return (total_time - preprocess_offset * nrepeat) / nrepeat;
#endif
}
else
......
library/include/ck/library/utility/check_err.hpp → include/ck/library/utility/check_err.hpp
View file @ 6e3c786e
......@@ -23,6 +23,130 @@
namespace ck {
namespace utils {
template <typename ComputeDataType, typename OutDataType, typename AccDataType = ComputeDataType>
double get_relative_threshold(const int number_of_accumulations = 1)
{
using F8 = ck::f8_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
static_assert(is_same_v<ComputeDataType, F8> || is_same_v<ComputeDataType, F16> ||
is_same_v<ComputeDataType, BF16> || is_same_v<ComputeDataType, F32> ||
is_same_v<ComputeDataType, I8> || is_same_v<ComputeDataType, I32> ||
is_same_v<ComputeDataType, int>,
"Warning: Unhandled ComputeDataType for setting up the relative threshold!");
double compute_error = 0;
if constexpr(is_same_v<ComputeDataType, I8> || is_same_v<ComputeDataType, I32> ||
is_same_v<ComputeDataType, int>)
{
return 0;
}
else
{
compute_error = std::pow(2, -NumericUtils<ComputeDataType>::mant) * 0.5;
}
static_assert(is_same_v<OutDataType, F8> || is_same_v<OutDataType, F16> ||
is_same_v<OutDataType, BF16> || is_same_v<OutDataType, F32> ||
is_same_v<OutDataType, I8> || is_same_v<OutDataType, I32> ||
is_same_v<OutDataType, int>,
"Warning: Unhandled OutDataType for setting up the relative threshold!");
double output_error = 0;
if constexpr(is_same_v<OutDataType, I8> || is_same_v<OutDataType, I32> ||
is_same_v<OutDataType, int>)
{
return 0;
}
else
{
output_error = std::pow(2, -NumericUtils<OutDataType>::mant) * 0.5;
}
double midway_error = std::max(compute_error, output_error);
static_assert(is_same_v<AccDataType, F8> || is_same_v<AccDataType, F16> ||
is_same_v<AccDataType, BF16> || is_same_v<AccDataType, F32> ||
is_same_v<AccDataType, I8> || is_same_v<AccDataType, I32> ||
is_same_v<AccDataType, int>,
"Warning: Unhandled AccDataType for setting up the relative threshold!");
double acc_error = 0;
if constexpr(is_same_v<AccDataType, I8> || is_same_v<AccDataType, I32> ||
is_same_v<AccDataType, int>)
{
return 0;
}
else
{
acc_error = std::pow(2, -NumericUtils<AccDataType>::mant) * 0.5 * number_of_accumulations;
}
return std::max(acc_error, midway_error);
}
template <typename ComputeDataType, typename OutDataType, typename AccDataType = ComputeDataType>
double get_absolute_threshold(const double max_possible_num, const int number_of_accumulations = 1)
{
using F8 = ck::f8_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
static_assert(is_same_v<ComputeDataType, F8> || is_same_v<ComputeDataType, F16> ||
is_same_v<ComputeDataType, BF16> || is_same_v<ComputeDataType, F32> ||
is_same_v<ComputeDataType, I8> || is_same_v<ComputeDataType, I32> ||
is_same_v<ComputeDataType, int>,
"Warning: Unhandled ComputeDataType for setting up the absolute threshold!");
auto expo = std::log2(std::abs(max_possible_num));
double compute_error = 0;
if constexpr(is_same_v<ComputeDataType, I8> || is_same_v<ComputeDataType, I32> ||
is_same_v<ComputeDataType, int>)
{
return 0;
}
else
{
compute_error = std::pow(2, expo - NumericUtils<ComputeDataType>::mant) * 0.5;
}
static_assert(is_same_v<OutDataType, F8> || is_same_v<OutDataType, F16> ||
is_same_v<OutDataType, BF16> || is_same_v<OutDataType, F32> ||
is_same_v<OutDataType, I8> || is_same_v<OutDataType, I32> ||
is_same_v<OutDataType, int>,
"Warning: Unhandled OutDataType for setting up the absolute threshold!");
double output_error = 0;
if constexpr(is_same_v<OutDataType, I8> || is_same_v<OutDataType, I32> ||
is_same_v<OutDataType, int>)
{
return 0;
}
else
{
output_error = std::pow(2, expo - NumericUtils<OutDataType>::mant) * 0.5;
}
double midway_error = std::max(compute_error, output_error);
static_assert(is_same_v<AccDataType, F8> || is_same_v<AccDataType, F16> ||
is_same_v<AccDataType, BF16> || is_same_v<AccDataType, F32> ||
is_same_v<AccDataType, I8> || is_same_v<AccDataType, I32> ||
is_same_v<AccDataType, int>,
"Warning: Unhandled AccDataType for setting up the absolute threshold!");
double acc_error = 0;
if constexpr(is_same_v<AccDataType, I8> || is_same_v<AccDataType, I32> ||
is_same_v<AccDataType, int>)
{
return 0;
}
else
{
acc_error =
std::pow(2, expo - NumericUtils<AccDataType>::mant) * 0.5 * number_of_accumulations;
}
return std::max(acc_error, midway_error);
}
template <typename Range, typename RefRange>
typename std::enable_if<
std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
......@@ -82,7 +206,7 @@ typename std::enable_if<
check_err(const Range& out,
const RefRange& ref,
const std::string& msg = "Error: Incorrect results!",
double rtol = 1e-3,
double rtol = 1e-1,
double atol = 1e-3)
{
if(out.size() != ref.size())
......@@ -253,11 +377,13 @@ check_err(const Range& out,
int err_count = 0;
double err = 0;
double max_err = std::numeric_limits<float>::min();
for(std::size_t i = 0; i < ref.size(); ++i)
{
const double o = type_convert<float>(*std::next(std::begin(out), i));
const double r = type_convert<float>(*std::next(std::begin(ref), i));
err = std::abs(o - r);
if(err > atol + rtol * std::abs(r) || !std::isfinite(o) || !std::isfinite(r))
{
max_err = err > max_err ? err : max_err;
......@@ -270,6 +396,7 @@ check_err(const Range& out,
res = false;
}
}
if(!res)
{
std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err
......
library/include/ck/library/utility/host_tensor.hpp → include/ck/library/utility/host_tensor.hpp
View file @ 6e3c786e
......@@ -332,7 +332,7 @@ struct Tensor
std::size_t GetElementSpaceSizeInBytes() const { return sizeof(T) * GetElementSpaceSize(); }
void SetZero() { ck::ranges::fill<T>(mData, 0); }
void SetZero() { ck::ranges::fill<T>(mData, T{0}); }
template <typename F>
void ForEach_impl(F&& f, std::vector<size_t>& idx, size_t rank)
......
library/include/ck/library/utility/host_tensor_generator.hpp → include/ck/library/utility/host_tensor_generator.hpp
View file @ 6e3c786e
// 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
......@@ -37,7 +37,7 @@ struct GeneratorTensor_1<ck::half_t>
float value = 1.0;
template <typename... Is>
ck::bhalf_t operator()(Is...)
ck::half_t operator()(Is...)
{
return ck::type_convert<ck::half_t>(value);
}
......@@ -62,7 +62,7 @@ struct GeneratorTensor_1<ck::f8_t>
float value = 1.0;
template <typename... Is>
ck::bhalf_t operator()(Is...)
ck::f8_t operator()(Is...)
{
return ck::type_convert<ck::f8_t>(value);
}
......@@ -286,14 +286,33 @@ struct GeneratorTensor_Checkboard
}
};
template <ck::index_t Dim>
/**
* @brief Is used to generate sequential values based on the specified dimension.
*
* @tparam T The type of the tensor values.
* @tparam Dim The specific dimension used for generation.
*
* GeneratorTensor_Sequential<1>{} will generate the following values for a 3x3 tensor:
*
* 0 1 2
* 0 1 2
* 0 1 2
*
* Essentially, the values generated are logical coordinates of the generated element that
* correspond to dimension Dim. E.g. for 2-dimensional tensor and Dim=1, the values are the column
* indices.
*
*/
template <typename T, ck::index_t Dim>
struct GeneratorTensor_Sequential
{
template <typename... Ts>
float operator()(Ts... Xs) const
T operator()(Ts... Xs) const
{
std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
return dims[Dim];
float tmp = dims[Dim];
return ck::type_convert<T>(tmp);
}
};
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.hpp
View file @ 6e3c786e
......@@ -269,15 +269,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
......@@ -341,14 +340,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
......@@ -396,14 +395,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
......@@ -447,14 +446,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
......@@ -760,15 +759,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
__builtin_amdgcn_sched_barrier(0);
// NOTE: Synchronize threads in a workgroup at the start of each MAC
......@@ -866,14 +864,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
__builtin_amdgcn_sched_barrier(0);
......@@ -942,14 +940,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
__builtin_amdgcn_sched_barrier(0);
......@@ -1018,14 +1016,14 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
__builtin_amdgcn_sched_barrier(0);
......
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