Commit 1b616990 authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCm/composable_kernel into update_cka8w8_uc

parents af30d6b6 800cf897
#!/bin/sh
EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
for fquant in "" "-fquant=1 -prec_o=int8" "-fquant=2 -prec_o=int8" "-fquant=1 -prec_o=fp8" "-fquant=2 -prec_o=fp8"; do
for pr_i in "fp16" "bf16" ; do
$EXE -prec=$pr_i -m=99 -n=13
$EXE -prec=$pr_i -m=17 -n=16
$EXE -prec=$pr_i -m=1 -n=100
$EXE -prec=$pr_i -m=4 -n=128
$EXE -prec=$pr_i -m=80 -n=127
$EXE -prec=$pr_i -m=22 -n=255 -stride=256
$EXE -prec=$pr_i -m=7 -n=599
$EXE -prec=$pr_i -m=19 -n=512
$EXE -prec=$pr_i -m=33 -n=313 -stride=1000
$EXE -prec=$pr_i -m=11 -n=510
$EXE -prec=$pr_i -m=171 -n=676 -stride=818
$EXE -prec=$pr_i -m=91 -n=636
$EXE -prec=$pr_i -m=12 -n=768 -stride=800
$EXE -prec=$pr_i -m=100 -n=766 -stride=812
$EXE -prec=$pr_i -m=31 -n=1024
$EXE -prec=$pr_i -m=64 -n=1000 -stride=1004
$EXE -prec=$pr_i -m=8 -n=1501
$EXE -prec=$pr_i -m=3 -n=1826
$EXE -prec=$pr_i -m=5 -n=2040
$EXE -prec=$pr_i -m=7 -n=2734
$EXE -prec=$pr_i -m=1 -n=3182
$EXE -prec=$pr_i -m=9 -n=4096
$EXE -prec=$pr_i -m=3 -n=8192
$EXE -prec=$pr_i -m=1 -n=10547
$EXE -prec=$pr_i -m=3 -n=17134
for fadd in "0" "1"; do
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=99 -n=13
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=17 -n=16
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1 -n=100
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=4 -n=128
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=80 -n=127
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=22 -n=255 -stride=256
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7 -n=599
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=19 -n=512
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=33 -n=313 -stride=1000
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=11 -n=510
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=171 -n=676 -stride=818
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=91 -n=636
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=12 -n=768 -stride=800
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=100 -n=766 -stride=812
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=31 -n=1024
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=64 -n=1000 -stride=1004
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=8 -n=1501
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3 -n=1826
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=5 -n=2040
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7 -n=2734
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1 -n=3182
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=9 -n=4096
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3 -n=8192
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1 -n=10547
#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3 -n=17134
done
done
done
......@@ -63,17 +63,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
assert(stride >= n);
assert(x_stride >= n);
using XDataType = DataType;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = DataType;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<SmoothScaleDataType> smscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
......@@ -82,15 +82,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
ck_tile::FillUniformDistribution<SmoothScaleDataType>{1e-3, .5f}(smscale_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem smscale_buf(smscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
xscale_buf.ToDevice(xscale_host.data());
smscale_buf.ToDevice(smscale_host.data());
constexpr bool kTwoPass = true;
......@@ -101,7 +101,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using Problem = ck_tile::SmoothquantPipelineProblem<XDataType,
XScaleDataType,
SmoothScaleDataType,
ComputeDataType,
YScaleDataType,
QYDataType,
......@@ -115,7 +115,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
using Kernel = ck_tile::Smoothquant<Pipeline>;
ck_tile::SmoothquantHostArgs args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
smscale_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
m,
......@@ -142,16 +142,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
// smooth outlier
{
auto f = [&](auto n_) {
auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
auto v_smscale = ck_tile::type_convert<ComputeDataType>(smscale_host(n_));
for(int m_ = 0; m_ < m; ++m_)
{
auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(m_, n_));
y_host(m_, n_) = v_x * v_xscale;
y_host(m_, n_) = v_x * v_smscale;
}
};
ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
ck_tile::make_ParallelTensorFunctor(f, smscale_host.get_element_space_size())(
std::thread::hardware_concurrency());
}
......
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "smoothquant.hpp"
......@@ -35,7 +35,7 @@ float smoothquant_(const S& s, A a)
using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
typename SmoothquantTypeConfig<DataType>::XDataType,
typename SmoothquantTypeConfig<DataType>::XScaleDataType,
typename SmoothquantTypeConfig<DataType>::SmoothScaleDataType,
typename SmoothquantTypeConfig<DataType>::ComputeDataType,
typename SmoothquantTypeConfig<DataType>::YScaleDataType,
typename SmoothquantTypeConfig<DataType>::QYDataType,
......
......@@ -66,15 +66,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
using TypeConfig = SmoothquantTypeConfig<DataType>;
using XDataType = typename TypeConfig::XDataType;
using XScaleDataType = typename TypeConfig::XScaleDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = typename TypeConfig::ComputeDataType;
using XDataType = typename TypeConfig::XDataType;
using SmoothScaleDataType = typename TypeConfig::SmoothScaleDataType;
using YScaleDataType = typename TypeConfig::YScaleDataType;
using QYDataType = typename TypeConfig::QYDataType;
using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<SmoothScaleDataType> smscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
......@@ -83,15 +83,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
ck_tile::FillUniformDistribution<SmoothScaleDataType>{1e-3, .5f}(smscale_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem smscale_buf(smscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
xscale_buf.ToDevice(xscale_host.data());
smscale_buf.ToDevice(smscale_host.data());
std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", x_stride:" << x_stride << ", y_stride:" << y_stride
......@@ -100,7 +100,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
smoothquant_traits traits{data_type};
smoothquant_args args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
smscale_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
m,
......@@ -111,7 +111,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
float ave_time = smoothquant(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(XScaleDataType) * n +
std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(SmoothScaleDataType) * n +
sizeof(YScaleDataType) * m + sizeof(QYDataType) * m * n;
float gb_per_sec = num_byte / 1.E6 / ave_time;
......@@ -126,16 +126,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
// smooth outlier
{
auto f = [&](auto n_) {
auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
auto v_smscale = ck_tile::type_convert<ComputeDataType>(smscale_host(n_));
for(int m_ = 0; m_ < m; ++m_)
{
auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(m_, n_));
y_host(m_, n_) = v_x * v_xscale;
y_host(m_, n_) = v_x * v_smscale;
}
};
ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
ck_tile::make_ParallelTensorFunctor(f, smscale_host.get_element_space_size())(
std::thread::hardware_concurrency());
}
......
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -14,21 +14,21 @@ struct SmoothquantTypeConfig;
template <>
struct SmoothquantTypeConfig<ck_tile::half_t>
{
using XDataType = ck_tile::half_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = ck_tile::half_t;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
template <>
struct SmoothquantTypeConfig<ck_tile::bf16_t>
{
using XDataType = ck_tile::bf16_t;
using XScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
using XDataType = ck_tile::bf16_t;
using SmoothScaleDataType = float;
using YScaleDataType = float;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
};
// runtime args
......
......@@ -15,8 +15,13 @@ template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true, false>>(const S&, A);
#endif
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,8 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,9 +6,14 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,7 +6,11 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 6, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 1024, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 256, 8, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 256, 4, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 1024, 2, true, true>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 1, true, true>>(const S&, A);
// clang-format on
......@@ -6,8 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 8, 4, 64, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,7 +6,11 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 1, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 1, 4, 64, 2, true , false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 4, 64, 1, true , false>>(const S&, A);
// clang-format on
......@@ -6,7 +6,11 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 3, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 6, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 12, 4, 64, 1, true, false>>(const S&, A);
// clang-format on
......@@ -15,8 +15,13 @@ template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true ,false>>(const S&, A);
#endif
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 2, 128, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 2, 128, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,8 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 4, 64, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 2, 128, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 6, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 1, 256, 8, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 1, 256, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 256, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 256, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,7 +6,11 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 1, 4, 64, 4, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 4, 64, 2, true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 4, 64, 1, true, false>>(const S&, A);
// clang-format on
......@@ -6,9 +6,13 @@
// clang-format off
// rm rn tm tn vn pd 2p
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 128, 8,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 4,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 2,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 3, 1, 1024, 1,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 1, 128, 8,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 1, 256, 4,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 6, 1, 256, 2,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 3, 1, 1024, 1,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 1, 128, 8,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 1, 256, 4,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 6, 1, 256, 2,true, false>>(const S&, A);
template float moe_smoothquant_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 3, 1, 1024, 1,true, 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