Skip to content

GitLab

Unverified Commit 8b49f207 authored by Max Podkorytov's avatar Max Podkorytov Committed by GitHub
Browse files

Merge branch 'develop' into fa-h512

parents 0d59f474 a6b761c3
Expand all Show whitespace changes
Inline Side-by-side
Showing with 2184 additions and 167 deletions
example/ck_tile/03_gemm/gemm_basic.hpp
View file @ 8b49f207
...@@ -51,25 +51,10 @@ using BDataType = Types::BDataType; ...@@ -51,25 +51,10 @@ using BDataType = Types::BDataType;
using AccDataType = Types::AccDataType; using AccDataType = Types::AccDataType;
using CDataType = Types::CDataType; using CDataType = Types::CDataType;
struct gemm_basic_args
{
const void* p_a;
const void* p_b;
void* p_c;
ck_tile::index_t kbatch;
ck_tile::index_t M;
ck_tile::index_t N;
ck_tile::index_t K;
ck_tile::index_t stride_A;
ck_tile::index_t stride_B;
ck_tile::index_t stride_C;
};
auto create_args(int argc, char* argv[]) auto create_args(int argc, char* argv[])
{ {
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
arg_parser.insert("b", "1", "batch size") arg_parser.insert("m", "3840", "m dimension")
.insert("m", "3840", "m dimension")
.insert("n", "4096", "n dimension") .insert("n", "4096", "n dimension")
.insert("k", "2048", "k dimension") .insert("k", "2048", "k dimension")
.insert("a_layout", "R", "A tensor data layout - Row by default") .insert("a_layout", "R", "A tensor data layout - Row by default")
...@@ -82,11 +67,12 @@ auto create_args(int argc, char* argv[]) ...@@ -82,11 +67,12 @@ auto create_args(int argc, char* argv[])
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8") .insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("warmup", "50", "number of iterations before benchmark the kernel") .insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel") .insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer"); .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("split_k", "1", "splitK value");
bool result = arg_parser.parse(argc, argv); bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser); return std::make_tuple(result, arg_parser);
} }
// host API // host API
float gemm_calc(gemm_basic_args args, const ck_tile::stream_config& s); float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s);
example/ck_tile/03_gemm/run_gemm_example.inc
View file @ 8b49f207
...@@ -16,11 +16,11 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf, ...@@ -16,11 +16,11 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
int n_warmup, int n_warmup,
int n_repeat) int n_repeat)
{ {
gemm_basic_args args; ck_tile::GemmHostArgs args;
args.p_a = a_m_k_dev_buf.GetDeviceBuffer(); args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
args.p_b = b_k_n_dev_buf.GetDeviceBuffer(); args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
args.p_c = c_m_n_dev_buf.GetDeviceBuffer(); args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
args.kbatch = kbatch; args.k_batch = kbatch;
args.M = M; args.M = M;
args.N = N; args.N = N;
args.K = K; args.K = K;
...@@ -64,7 +64,7 @@ int run_gemm_example_with_layouts(int argc, ...@@ -64,7 +64,7 @@ int run_gemm_example_with_layouts(int argc,
ck_tile::index_t stride_B = arg_parser.get_int("stride_b"); ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
ck_tile::index_t stride_C = arg_parser.get_int("stride_c"); ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
ck_tile::index_t batch_size = arg_parser.get_int("b"); ck_tile::index_t kbatch = arg_parser.get_int("split_k");
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");
...@@ -133,7 +133,7 @@ int run_gemm_example_with_layouts(int argc, ...@@ -133,7 +133,7 @@ int run_gemm_example_with_layouts(int argc,
stride_A, stride_A,
stride_B, stride_B,
stride_C, stride_C,
batch_size, kbatch,
n_warmup, n_warmup,
n_repeat); n_repeat);
...@@ -161,14 +161,39 @@ int run_gemm_example_with_layouts(int argc, ...@@ -161,14 +161,39 @@ int run_gemm_example_with_layouts(int argc,
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();
ADataType* d_A;
BDataType* d_B;
CDataType* d_C;
ck_tile::hip_check_error(hipMalloc(&d_A, M * K * sizeof(ADataType)));
ck_tile::hip_check_error(hipMalloc(&d_B, N * K * sizeof(BDataType)));
ck_tile::hip_check_error(hipMalloc(&d_C, M * N * sizeof(CDataType)));
ck_tile::hip_check_error(hipMemcpy(d_A,
a_m_k_dev_buf.GetDeviceBuffer(),
M * K * sizeof(ADataType),
hipMemcpyHostToDevice));
ck_tile::hip_check_error(hipMemcpy(d_B,
b_k_n_dev_buf.GetDeviceBuffer(),
N * K * sizeof(BDataType),
hipMemcpyHostToDevice));
ck_tile::reference_gemm_gpu<ADataType, ck_tile::reference_gemm_gpu<ADataType,
BDataType, BDataType,
AccDataType, AccDataType,
CDataType, CDataType,
ALayout, ALayout,
BLayout, BLayout,
CLayout>( CLayout>(d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
a_m_k_dev_buf, b_k_n_dev_buf, c_m_n_gpu_buf_ref, M, N, K, stride_A, stride_B, stride_C);
ck_tile::hip_check_error(hipMemcpy(c_m_n_gpu_buf_ref.GetDeviceBuffer(),
d_C,
M * N * sizeof(CDataType),
hipMemcpyDeviceToHost));
ck_tile::hip_check_error(hipFree(d_A));
ck_tile::hip_check_error(hipFree(d_B));
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); pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_gpu_ref);
......
example/ck_tile/03_gemm/universal_gemm.cpp
View file @ 8b49f207
...@@ -22,7 +22,7 @@ ...@@ -22,7 +22,7 @@
#endif #endif
template <typename ALayout, typename BLayout, typename CLayout> template <typename ALayout, typename BLayout, typename CLayout>
float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s) float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{ {
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY) #if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
// Memory friendly for Interwave scheduler // Memory friendly for Interwave scheduler
...@@ -78,7 +78,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s) ...@@ -78,7 +78,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
#endif #endif
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>; ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(args.K); const ck_tile::index_t k_grain = args.k_batch * K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop); const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop); const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
...@@ -106,17 +108,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s) ...@@ -106,17 +108,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
has_hot_loop_v, has_hot_loop_v,
tail_number_v>>; tail_number_v>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>; using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args.p_a, auto kargs = Kernel::MakeKernelArgs(args);
args.p_b,
args.p_c, const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
args.M,
args.N,
args.K,
args.stride_A,
args.stride_B,
args.stride_C);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize(); constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs)) if(!Kernel::IsSupportedArgument(kargs))
......
example/ck_tile/12_smoothquant/example_smoothquant.cpp
View file @ 8b49f207
...@@ -35,7 +35,8 @@ auto create_args(int argc, char* argv[]) ...@@ -35,7 +35,8 @@ auto create_args(int argc, char* argv[])
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension") arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension") .insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n") .insert("x_stride", "-1", "input stride per row, if -1 then equal to n")
.insert("y_stride", "-1", "output stride per row, if -1 then equal to n")
.insert("e", "1e-5", "epsilon") .insert("e", "1e-5", "epsilon")
.insert("v", "1", "cpu validation or not") .insert("v", "1", "cpu validation or not")
.insert("prec", "fp16", "precision") .insert("prec", "fp16", "precision")
...@@ -51,9 +52,12 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -51,9 +52,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
{ {
ck_tile::index_t m = arg_parser.get_int("m"); ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n"); ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride"); ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
if(stride < 0) if(x_stride < 0)
stride = n; x_stride = n;
ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
if(y_stride < 0)
y_stride = n;
std::string data_type = arg_parser.get_str("prec"); std::string data_type = arg_parser.get_str("prec");
int do_validation = arg_parser.get_int("v"); int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup"); int warmup = arg_parser.get_int("warmup");
...@@ -68,14 +72,14 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -68,14 +72,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
using ComputeDataType = float; using ComputeDataType = float;
// host verify // host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1}); ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n}); ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1}); ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1}); ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1}); ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {y_stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1}); ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host); ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host); ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
...@@ -116,7 +120,8 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -116,7 +120,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.GetDeviceBuffer(), qy_buf.GetDeviceBuffer(),
m, m,
n, n,
stride}; x_stride,
y_stride};
auto kargs = Kernel::MakeKargs(args); auto kargs = Kernel::MakeKargs(args);
...@@ -133,7 +138,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -133,7 +138,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(do_validation) if(do_validation)
{ {
using YDataType = ComputeDataType; using YDataType = ComputeDataType;
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {stride, 1}); ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {y_stride, 1});
// smooth outlier // smooth outlier
{ {
auto f = [&](auto n_) { auto f = [&](auto n_) {
...@@ -183,7 +188,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -183,7 +188,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.FromDevice(qy_host_dev.data()); qy_buf.FromDevice(qy_host_dev.data());
auto [rtol, atol] = get_elimit<QYDataType>(); auto [rtol, atol] = get_elimit<QYDataType>();
if(stride == n) if(y_stride == n)
{ {
pass = ck_tile::check_err(qy_host_dev, pass = ck_tile::check_err(qy_host_dev,
qy_host_ref, qy_host_ref,
...@@ -195,10 +200,12 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -195,10 +200,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
{ {
for(int i_r = 0; i_r < m; i_r++) for(int i_r = 0; i_r < m; i_r++)
{ {
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride, std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * y_stride,
qy_host_dev.begin() + i_r * stride + n); qy_host_dev.begin() + i_r * y_stride +
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride, n);
qy_host_ref.begin() + i_r * stride + n); std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * y_stride,
qy_host_ref.begin() + i_r * y_stride +
n);
pass &= ck_tile::check_err(qy_host_dev_row, pass &= ck_tile::check_err(qy_host_dev_row,
qy_host_ref_row, qy_host_ref_row,
std::string("qy[") + std::to_string(i_r) + std::string("qy[") + std::to_string(i_r) +
...@@ -210,8 +217,9 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -210,8 +217,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
} }
std::cout << "[" << data_type << "]" std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride << " m:" << m << ", n:" << n << ", x_stride:" << x_stride
<< ", valid:" << (pass ? "y" : "n") << std::flush << std::endl; << ", y_stride:" << y_stride << ", valid:" << (pass ? "y" : "n") << std::flush
<< std::endl;
} }
return pass; return pass;
......
example/ck_tile/12_smoothquant/smoothquant.cpp
View file @ 8b49f207
...@@ -33,7 +33,8 @@ auto create_args(int argc, char* argv[]) ...@@ -33,7 +33,8 @@ auto create_args(int argc, char* argv[])
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension") arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension") .insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n") .insert("x_stride", "-1", "input stride per row, if -1 then equal to n")
.insert("y_stride", "-1", "output stride per row, if -1 then equal to n")
.insert("v", "1", "cpu validation or not") .insert("v", "1", "cpu validation or not")
.insert("kname", "1", "print kernel name or not") .insert("kname", "1", "print kernel name or not")
.insert("prec", "fp16", "precision") .insert("prec", "fp16", "precision")
...@@ -49,16 +50,19 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -49,16 +50,19 @@ bool run(const ck_tile::ArgParser& arg_parser)
{ {
ck_tile::index_t m = arg_parser.get_int("m"); ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n"); ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride"); ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
if(stride < 0) if(x_stride < 0)
stride = n; x_stride = n;
ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
if(y_stride < 0)
y_stride = n;
std::string data_type = arg_parser.get_str("prec"); std::string data_type = arg_parser.get_str("prec");
int kname = arg_parser.get_int("kname"); int kname = arg_parser.get_int("kname");
int do_validation = arg_parser.get_int("v"); int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup"); int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat"); int repeat = arg_parser.get_int("repeat");
assert(stride >= n); assert(x_stride >= n);
using TypeConfig = SmoothquantTypeConfig<DataType>; using TypeConfig = SmoothquantTypeConfig<DataType>;
...@@ -69,14 +73,14 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -69,14 +73,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
using ComputeDataType = typename TypeConfig::ComputeDataType; using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify // host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1}); ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n}); ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1}); ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1}); ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1}); ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {y_stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1}); ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host); ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host); ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
...@@ -90,7 +94,8 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -90,7 +94,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
xscale_buf.ToDevice(xscale_host.data()); xscale_buf.ToDevice(xscale_host.data());
std::cout << "[" << data_type << "]" std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride << std::flush; << " m:" << m << ", n:" << n << ", x_stride:" << x_stride << ", y_stride:" << y_stride
<< std::flush;
smoothquant_traits traits{data_type}; smoothquant_traits traits{data_type};
...@@ -100,7 +105,8 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -100,7 +105,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.GetDeviceBuffer(), qy_buf.GetDeviceBuffer(),
m, m,
n, n,
stride}; x_stride,
y_stride};
float ave_time = smoothquant( float ave_time = smoothquant(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat}); traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
...@@ -116,7 +122,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -116,7 +122,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(do_validation) if(do_validation)
{ {
using YDataType = ComputeDataType; using YDataType = ComputeDataType;
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {stride, 1}); ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {y_stride, 1});
// smooth outlier // smooth outlier
{ {
auto f = [&](auto n_) { auto f = [&](auto n_) {
...@@ -166,7 +172,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -166,7 +172,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
qy_buf.FromDevice(qy_host_dev.data()); qy_buf.FromDevice(qy_host_dev.data());
auto [rtol, atol] = get_elimit<QYDataType>(); auto [rtol, atol] = get_elimit<QYDataType>();
if(stride == n) if(y_stride == n)
{ {
pass = ck_tile::check_err(qy_host_dev, pass = ck_tile::check_err(qy_host_dev,
qy_host_ref, qy_host_ref,
...@@ -178,10 +184,12 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -178,10 +184,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
{ {
for(int i_r = 0; i_r < m; i_r++) for(int i_r = 0; i_r < m; i_r++)
{ {
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride, std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * y_stride,
qy_host_dev.begin() + i_r * stride + n); qy_host_dev.begin() + i_r * y_stride +
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride, n);
qy_host_ref.begin() + i_r * stride + n); std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * y_stride,
qy_host_ref.begin() + i_r * y_stride +
n);
pass &= ck_tile::check_err(qy_host_dev_row, pass &= ck_tile::check_err(qy_host_dev_row,
qy_host_ref_row, qy_host_ref_row,
std::string("qy[") + std::to_string(i_r) + std::string("qy[") + std::to_string(i_r) +
......
example/ck_tile/13_moe_sorting/moe_sorting_api.cpp
View file @ 8b49f207
...@@ -3,9 +3,11 @@ ...@@ -3,9 +3,11 @@
#include "moe_sorting_api.hpp" #include "moe_sorting_api.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \ #define MOE_SORTING_DISPATCH_ETILE(unroll_num_, expert_tile_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \ constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \ constexpr ck_tile::index_t expert_tile = expert_tile_; \
using ms_problem = \
ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num, expert_tile>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \ using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \ auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \ const dim3 grids = kernel::GridSize(a); \
...@@ -15,6 +17,28 @@ ...@@ -15,6 +17,28 @@
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \ s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time; return ave_time;
#define MOE_SORTING_DISPATCH(unroll_num_) \
if(a.num_experts <= 8) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 8) \
} \
else if(a.num_experts <= 16) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 16) \
} \
else if(a.num_experts <= 32) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 32) \
} \
else if(a.num_experts <= 64) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 64) \
} \
else \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 0) \
}
float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s) float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s)
{ {
if(t.weight_type == "fp32" && t.index_type == "int32") if(t.weight_type == "fp32" && t.index_type == "int32")
...@@ -49,21 +73,12 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi ...@@ -49,21 +73,12 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi
case(6): { case(6): {
MOE_SORTING_DISPATCH(6); MOE_SORTING_DISPATCH(6);
} }
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): { case(8): {
MOE_SORTING_DISPATCH(8); MOE_SORTING_DISPATCH(8);
} }
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): { case(10): {
MOE_SORTING_DISPATCH(10); MOE_SORTING_DISPATCH(10);
} }
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: { default: {
MOE_SORTING_DISPATCH(4); MOE_SORTING_DISPATCH(4);
} }
......
example/ck_tile/13_moe_sorting/script/smoke_test.sh
View file @ 8b49f207
...@@ -17,3 +17,4 @@ $EXE -t=71 -e=11 -k=11 ...@@ -17,3 +17,4 @@ $EXE -t=71 -e=11 -k=11
$EXE -t=1 -e=1 -k=1 $EXE -t=1 -e=1 -k=1
$EXE -t=99 -e=2 -k=1 $EXE -t=99 -e=2 -k=1
$EXE -t=333 -e=99 -k=13 $EXE -t=333 -e=99 -k=13
$EXE -t=128 -e=32 -k=5 -moe_buf_size=262144
example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp
View file @ 8b49f207
...@@ -3,9 +3,11 @@ ...@@ -3,9 +3,11 @@
#include "fused_moesorting.hpp" #include "fused_moesorting.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \ #define MOE_SORTING_DISPATCH_ETILE(unroll_num_, expert_tile_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \ constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \ constexpr ck_tile::index_t expert_tile = expert_tile_; \
using ms_problem = \
ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num, expert_tile>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \ using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \ auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \ const dim3 grids = kernel::GridSize(a); \
...@@ -15,6 +17,28 @@ ...@@ -15,6 +17,28 @@
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \ s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time; return ave_time;
#define MOE_SORTING_DISPATCH(unroll_num_) \
if(a.num_experts <= 8) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 8) \
} \
else if(a.num_experts <= 16) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 16) \
} \
else if(a.num_experts <= 32) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 32) \
} \
else if(a.num_experts <= 64) \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 64) \
} \
else \
{ \
MOE_SORTING_DISPATCH_ETILE(unroll_num_, 0) \
}
float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s) float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s)
{ {
if(t.weight_type == "fp32" && t.index_type == "int32") if(t.weight_type == "fp32" && t.index_type == "int32")
...@@ -49,21 +73,12 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til ...@@ -49,21 +73,12 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til
case(6): { case(6): {
MOE_SORTING_DISPATCH(6); MOE_SORTING_DISPATCH(6);
} }
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): { case(8): {
MOE_SORTING_DISPATCH(8); MOE_SORTING_DISPATCH(8);
} }
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): { case(10): {
MOE_SORTING_DISPATCH(10); MOE_SORTING_DISPATCH(10);
} }
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: { default: {
MOE_SORTING_DISPATCH(4); MOE_SORTING_DISPATCH(4);
} }
......
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View file @ 8b49f207
...@@ -16,7 +16,7 @@ ...@@ -16,7 +16,7 @@
#include "batched_gemm.hpp" #include "batched_gemm.hpp"
template <typename ALayout, typename BLayout, typename CLayout> template <typename ALayout, typename BLayout, typename CLayout>
float batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config& s) float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stream_config& s)
{ {
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part. // The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false; constexpr bool kPadM = false;
...@@ -70,20 +70,25 @@ float batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config& ...@@ -70,20 +70,25 @@ float batched_gemm(const batched_gemm_kargs& args, const ck_tile::stream_config&
using CodegenGemmTraits = using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>; ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile:: using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>; GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>; using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM. // ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy. // Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>; using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args); auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args); const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch, args.batch_count);
constexpr dim3 blocks = Kernel::BlockSize(); constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0) if(s.log_level_ > 0)
{ {
std::cout << "Launching kernel with args:" std::cout << "Launching kernel with args:"
......
example/ck_tile/16_batched_gemm/batched_gemm.hpp
View file @ 8b49f207
...@@ -29,10 +29,6 @@ using BDataType = Types::BDataType; ...@@ -29,10 +29,6 @@ using BDataType = Types::BDataType;
using AccDataType = Types::AccDataType; using AccDataType = Types::AccDataType;
using CDataType = Types::CDataType; using CDataType = Types::CDataType;
struct batched_gemm_kargs : public ck_tile::BatchedGemmHostArgs
{
};
auto create_args(int argc, char* argv[]) auto create_args(int argc, char* argv[])
{ {
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
...@@ -53,11 +49,12 @@ auto create_args(int argc, char* argv[]) ...@@ -53,11 +49,12 @@ auto create_args(int argc, char* argv[])
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8") .insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("warmup", "50", "number of iterations before benchmark the kernel") .insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel") .insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer"); .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("split_k", "1", "splitK value");
bool result = arg_parser.parse(argc, argv); bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser); return std::make_tuple(result, arg_parser);
} }
// host API // host API
float batched_gemm(batched_gemm_kargs args, const ck_tile::stream_config& s); float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stream_config& s);
example/ck_tile/16_batched_gemm/run_batched_gemm_example.inc
View file @ 8b49f207
...@@ -17,13 +17,15 @@ float invoke_batched_gemm(ck_tile::DeviceMem& a_m_k_dev_buf, ...@@ -17,13 +17,15 @@ float invoke_batched_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::index_t batch_stride_B, ck_tile::index_t batch_stride_B,
ck_tile::index_t batch_stride_C, ck_tile::index_t batch_stride_C,
ck_tile::index_t batch_count, ck_tile::index_t batch_count,
ck_tile::index_t kbatch,
int n_warmup, int n_warmup,
int n_repeat) int n_repeat)
{ {
batched_gemm_kargs args; ck_tile::BatchedGemmHostArgs args;
args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer(); args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer(); args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer(); args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
args.k_batch = kbatch;
args.M = M; args.M = M;
args.N = N; args.N = N;
args.K = K; args.K = K;
...@@ -79,6 +81,7 @@ int run_batched_gemm_example_with_layouts(int argc, ...@@ -79,6 +81,7 @@ int run_batched_gemm_example_with_layouts(int argc,
ck_tile::index_t batch_stride_B = arg_parser.get_int("batch_stride_b"); ck_tile::index_t batch_stride_B = arg_parser.get_int("batch_stride_b");
ck_tile::index_t batch_stride_C = arg_parser.get_int("batch_stride_c"); ck_tile::index_t batch_stride_C = arg_parser.get_int("batch_stride_c");
ck_tile::index_t batch_count = arg_parser.get_int("batch_count"); ck_tile::index_t batch_count = arg_parser.get_int("batch_count");
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
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");
...@@ -159,6 +162,7 @@ int run_batched_gemm_example_with_layouts(int argc, ...@@ -159,6 +162,7 @@ int run_batched_gemm_example_with_layouts(int argc,
batch_stride_B, batch_stride_B,
batch_stride_C, batch_stride_C,
batch_count, batch_count,
kbatch,
n_warmup, n_warmup,
n_repeat); n_repeat);
...@@ -188,15 +192,33 @@ int run_batched_gemm_example_with_layouts(int argc, ...@@ -188,15 +192,33 @@ int run_batched_gemm_example_with_layouts(int argc,
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();
ADataType* d_A;
BDataType* d_B;
CDataType* d_C;
ck_tile::hip_check_error(hipMalloc(&d_A, batch_count * M * K * sizeof(ADataType)));
ck_tile::hip_check_error(hipMalloc(&d_B, batch_count * N * K * sizeof(BDataType)));
ck_tile::hip_check_error(hipMalloc(&d_C, batch_count * M * N * sizeof(CDataType)));
ck_tile::hip_check_error(hipMemcpy(d_A,
a_m_k_dev_buf.GetDeviceBuffer(),
batch_count * M * K * sizeof(ADataType),
hipMemcpyHostToDevice));
ck_tile::hip_check_error(hipMemcpy(d_B,
b_k_n_dev_buf.GetDeviceBuffer(),
batch_count * N * K * sizeof(BDataType),
hipMemcpyHostToDevice));
ck_tile::reference_batched_gemm_gpu<ADataType, ck_tile::reference_batched_gemm_gpu<ADataType,
BDataType, BDataType,
AccDataType, AccDataType,
CDataType, CDataType,
ALayout, ALayout,
BLayout, BLayout,
CLayout>(a_m_k_dev_buf, CLayout>(d_A,
b_k_n_dev_buf, d_B,
c_m_n_gpu_buf_ref, d_C,
M, M,
N, N,
K, K,
...@@ -208,6 +230,15 @@ int run_batched_gemm_example_with_layouts(int argc, ...@@ -208,6 +230,15 @@ int run_batched_gemm_example_with_layouts(int argc,
batch_stride_C, batch_stride_C,
batch_count); batch_count);
ck_tile::hip_check_error(hipMemcpy(c_m_n_gpu_buf_ref.GetDeviceBuffer(),
d_C,
batch_count * M * N * sizeof(CDataType),
hipMemcpyDeviceToHost));
ck_tile::hip_check_error(hipFree(d_A));
ck_tile::hip_check_error(hipFree(d_B));
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); pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_gpu_ref);
......
example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
View file @ 8b49f207
...@@ -34,13 +34,19 @@ using grouped_gemm_kargs = ck_tile::GroupedGemmHostArgs; ...@@ -34,13 +34,19 @@ using grouped_gemm_kargs = ck_tile::GroupedGemmHostArgs;
auto create_args(int argc, char* argv[]) auto create_args(int argc, char* argv[])
{ {
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
arg_parser.insert("a_layout", "R", "A tensor data layout - Row by default") arg_parser.insert("Ms", "", "M dimensions - empty by default.")
.insert("b_layout", "R", "B tensor data layout - Row by default") .insert("Ns", "", "N dimensions - empty by default.")
.insert("c_layout", "R", "C tensor data layout - Row by default") .insert("Ks", "", "K dimensions - empty by default.")
.insert("validate", "1", "0. No validation, 1. Validation on CPU") .insert("stride_As", "", "Tensor A strides - it is empty by default.")
.insert("warmup", "10", "number of iterations before benchmark the kernel") .insert("stride_Bs", "", "Tensor B strides - it is empty by default.")
.insert("repeat", "100", "number of iterations to benchmark the kernel") .insert("stride_Cs", "", "Tensor C strides - it is empty by default.")
.insert("group_count", "16", "group count"); .insert("a_layout", "R", "A tensor data layout - Row by default.")
.insert("b_layout", "R", "B tensor data layout - Row by default.")
.insert("c_layout", "R", "C tensor data layout - Row by default.")
.insert("validate", "1", "0. No validation, 1. Validation on CPU.")
.insert("warmup", "10", "number of iterations before benchmark the kernel.")
.insert("repeat", "100", "number of iterations to benchmark the kernel.")
.insert("group_count", "16", "group count.");
bool result = arg_parser.parse(argc, argv); bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser); return std::make_tuple(result, arg_parser);
......
example/ck_tile/17_grouped_gemm/run_grouped_gemm_example.inc
View file @ 8b49f207
...@@ -53,17 +53,24 @@ int run_grouped_gemm_example_with_layouts(int argc, ...@@ -53,17 +53,24 @@ int run_grouped_gemm_example_with_layouts(int argc,
return -1; return -1;
}; };
auto valid_input_data = [&](int group_count, const auto&... args) {
return !(args.empty() || ...) && group_count == (args.size() == ...);
};
const int group_count = arg_parser.get_int("group_count"); const int group_count = arg_parser.get_int("group_count");
const int repeat = arg_parser.get_int("repeat"); const int repeat = arg_parser.get_int("repeat");
const int warmup = arg_parser.get_int("warmup"); const int warmup = arg_parser.get_int("warmup");
std::vector<ck_tile::index_t> Ms; std::vector<ck_tile::index_t> Ms = arg_parser.get_int_vec("Ms");
std::vector<ck_tile::index_t> Ns; std::vector<ck_tile::index_t> Ns = arg_parser.get_int_vec("Ns");
std::vector<ck_tile::index_t> Ks; std::vector<ck_tile::index_t> Ks = arg_parser.get_int_vec("Ks");
std::vector<ck_tile::index_t> stride_As; std::vector<ck_tile::index_t> stride_As = arg_parser.get_int_vec("stride_As");
std::vector<ck_tile::index_t> stride_Bs; std::vector<ck_tile::index_t> stride_Bs = arg_parser.get_int_vec("stride_Bs");
std::vector<ck_tile::index_t> stride_Cs; std::vector<ck_tile::index_t> stride_Cs = arg_parser.get_int_vec("stride_Cs");
if(!valid_input_data(group_count, Ms, Ns, Ks, stride_As, stride_Bs, stride_Cs))
{
std::cout << "Please check the input data. Default values will be used." << std::endl;
for(int i = 0; i < group_count; i++) for(int i = 0; i < group_count; i++)
{ {
Ms.push_back(256 + 256 * i); Ms.push_back(256 + 256 * i);
...@@ -74,6 +81,7 @@ int run_grouped_gemm_example_with_layouts(int argc, ...@@ -74,6 +81,7 @@ int run_grouped_gemm_example_with_layouts(int argc,
stride_Bs.push_back(Ks[i]); stride_Bs.push_back(Ks[i]);
stride_Cs.push_back(Ns[i]); stride_Cs.push_back(Ns[i]);
} }
}
std::vector<ck_tile::HostTensor<ADataType>> a_m_k_tensors; std::vector<ck_tile::HostTensor<ADataType>> a_m_k_tensors;
std::vector<ck_tile::HostTensor<BDataType>> b_k_n_tensors; std::vector<ck_tile::HostTensor<BDataType>> b_k_n_tensors;
......
include/ck/config.h.in
View file @ 8b49f207
...@@ -111,6 +111,22 @@ ...@@ -111,6 +111,22 @@
#cmakedefine CK_USE_WMMA @CK_USE_WMMA@ #cmakedefine CK_USE_WMMA @CK_USE_WMMA@
#endif #endif
#ifndef CK_USE_GFX94
#cmakedefine CK_USE_GFX94 @CK_USE_GFX94@
#endif
#ifndef CK_USE_OCP_FP8
#cmakedefine CK_USE_OCP_FP8 @CK_USE_OCP_FP8@
#endif
#ifndef CK_USE_FNUZ_FP8
#cmakedefine CK_USE_FNUZ_FP8 @CK_USE_FNUZ_FP8@
#endif
#ifndef CK_USE_FP8_ON_UNSUPPORTED_ARCH
#cmakedefine CK_USE_FP8_ON_UNSUPPORTED_ARCH @CK_USE_FP8_ON_UNSUPPORTED_ARCH@
#endif
// clang-format on // clang-format on
#endif // CK_CONFIG_H_IN #endif // CK_CONFIG_H_IN
include/ck/library/utility/host_tensor.hpp
View file @ 8b49f207
// SPDX-License-Identifier: MIT // 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 #pragma once
...@@ -44,10 +44,19 @@ std::ostream& LogRangeAsType(std::ostream& os, Range&& range, std::string delim) ...@@ -44,10 +44,19 @@ std::ostream& LogRangeAsType(std::ostream& os, Range&& range, std::string delim)
else else
os << delim; os << delim;
if constexpr(std::is_same_v<T, ck::f8_t> || std::is_same_v<T, ck::bf8_t>) using RangeType = ck::remove_cvref_t<decltype(v)>;
if constexpr(std::is_same_v<RangeType, ck::f8_t> || std::is_same_v<RangeType, ck::bf8_t> ||
std::is_same_v<RangeType, ck::bhalf_t>)
{ {
os << ck::type_convert<float>(v); os << ck::type_convert<float>(v);
} }
else if constexpr(std::is_same_v<RangeType, ck::pk_i4_t>)
{
const auto packed_floats = ck::type_convert<ck::float2_t>(v);
const ck::vector_type<float, 2> vector_of_floats{packed_floats};
os << vector_of_floats.template AsType<float>()[ck::Number<0>{}] << delim
<< vector_of_floats.template AsType<float>()[ck::Number<1>{}];
}
else else
{ {
os << static_cast<T>(v); os << static_cast<T>(v);
...@@ -266,18 +275,18 @@ struct Tensor ...@@ -266,18 +275,18 @@ struct Tensor
using Data = std::vector<T>; using Data = std::vector<T>;
template <typename X> template <typename X>
Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize()) Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(GetElementSpaceSize())
{ {
} }
template <typename X, typename Y> template <typename X, typename Y>
Tensor(std::initializer_list<X> lens, std::initializer_list<Y> strides) Tensor(std::initializer_list<X> lens, std::initializer_list<Y> strides)
: mDesc(lens, strides), mData(mDesc.GetElementSpaceSize()) : mDesc(lens, strides), mData(GetElementSpaceSize())
{ {
} }
template <typename Lengths> template <typename Lengths>
Tensor(const Lengths& lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize()) Tensor(const Lengths& lens) : mDesc(lens), mData(GetElementSpaceSize())
{ {
} }
...@@ -287,7 +296,7 @@ struct Tensor ...@@ -287,7 +296,7 @@ struct Tensor
{ {
} }
Tensor(const Descriptor& desc) : mDesc(desc), mData(mDesc.GetElementSpaceSize()) {} Tensor(const Descriptor& desc) : mDesc(desc), mData(GetElementSpaceSize()) {}
template <typename OutT> template <typename OutT>
Tensor<OutT> CopyAsType() const Tensor<OutT> CopyAsType() const
...@@ -322,7 +331,17 @@ struct Tensor ...@@ -322,7 +331,17 @@ struct Tensor
std::size_t GetElementSize() const { return mDesc.GetElementSize(); } std::size_t GetElementSize() const { return mDesc.GetElementSize(); }
std::size_t GetElementSpaceSize() const { return mDesc.GetElementSpaceSize(); } std::size_t GetElementSpaceSize() const
{
if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
{
return (mDesc.GetElementSpaceSize() + 1) / 2;
}
else
{
return mDesc.GetElementSpaceSize();
}
}
std::size_t GetElementSpaceSizeInBytes() const { return sizeof(T) * GetElementSpaceSize(); } std::size_t GetElementSpaceSizeInBytes() const { return sizeof(T) * GetElementSpaceSize(); }
...@@ -468,31 +487,66 @@ struct Tensor ...@@ -468,31 +487,66 @@ struct Tensor
template <typename... Is> template <typename... Is>
std::size_t GetOffsetFromMultiIndex(Is... is) const std::size_t GetOffsetFromMultiIndex(Is... is) const
{
if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
{
return mDesc.GetOffsetFromMultiIndex(is...) / 2;
}
else
{ {
return mDesc.GetOffsetFromMultiIndex(is...); return mDesc.GetOffsetFromMultiIndex(is...);
} }
}
template <typename... Is> template <typename... Is>
T& operator()(Is... is) T& operator()(Is... is)
{
if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
{
return mData[mDesc.GetOffsetFromMultiIndex(is...) / 2];
}
else
{ {
return mData[mDesc.GetOffsetFromMultiIndex(is...)]; return mData[mDesc.GetOffsetFromMultiIndex(is...)];
} }
}
template <typename... Is> template <typename... Is>
const T& operator()(Is... is) const const T& operator()(Is... is) const
{
if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
{
return mData[mDesc.GetOffsetFromMultiIndex(is...) / 2];
}
else
{ {
return mData[mDesc.GetOffsetFromMultiIndex(is...)]; return mData[mDesc.GetOffsetFromMultiIndex(is...)];
} }
}
T& operator()(std::vector<std::size_t> idx) T& operator()(std::vector<std::size_t> idx)
{
if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
{
return mData[mDesc.GetOffsetFromMultiIndex(idx) / 2];
}
else
{ {
return mData[mDesc.GetOffsetFromMultiIndex(idx)]; return mData[mDesc.GetOffsetFromMultiIndex(idx)];
} }
}
const T& operator()(std::vector<std::size_t> idx) const const T& operator()(std::vector<std::size_t> idx) const
{
if constexpr(ck::is_same_v<ck::remove_cvref_t<T>, ck::pk_i4_t>)
{
return mData[mDesc.GetOffsetFromMultiIndex(idx) / 2];
}
else
{ {
return mData[mDesc.GetOffsetFromMultiIndex(idx)]; return mData[mDesc.GetOffsetFromMultiIndex(idx)];
} }
}
typename Data::iterator begin() { return mData.begin(); } typename Data::iterator begin() { return mData.begin(); }
......
include/ck/library/utility/host_tensor_generator.hpp
View file @ 8b49f207
...@@ -81,6 +81,20 @@ struct GeneratorTensor_1<int8_t> ...@@ -81,6 +81,20 @@ struct GeneratorTensor_1<int8_t>
} }
}; };
template <>
struct GeneratorTensor_1<ck::pk_i4_t>
{
int8_t value = 1;
template <typename... Is>
ck::pk_i4_t operator()(Is...)
{
int t = value + 8;
ck::pk_i4_t r = ((t << 4) + t) & 0xff;
return r;
}
};
template <typename T> template <typename T>
struct GeneratorTensor_2 struct GeneratorTensor_2
{ {
...@@ -121,6 +135,22 @@ struct GeneratorTensor_2<int8_t> ...@@ -121,6 +135,22 @@ struct GeneratorTensor_2<int8_t>
} }
}; };
template <>
struct GeneratorTensor_2<ck::pk_i4_t>
{
int min_value = 0;
int max_value = 1;
template <typename... Is>
ck::pk_i4_t operator()(Is...)
{
int hi = std::rand() % (max_value - min_value) + min_value + 8;
int lo = std::rand() % (max_value - min_value) + min_value + 8;
ck::pk_i4_t r = ((hi << 4) + lo) & 0xff;
return r;
}
};
#if defined CK_ENABLE_FP8 #if defined CK_ENABLE_FP8
template <> template <>
struct GeneratorTensor_2<ck::f8_t> struct GeneratorTensor_2<ck::f8_t>
......
include/ck/tensor/static_tensor.hpp
View file @ 8b49f207
...@@ -167,7 +167,7 @@ struct StaticTensorTupleOfVectorBuffer ...@@ -167,7 +167,7 @@ struct StaticTensorTupleOfVectorBuffer
// Idx is for S, not X. Idx should be aligned with X // Idx is for S, not X. Idx should be aligned with X
template <typename X, template <typename X,
typename Idx, typename Idx,
typename enable_if<has_same_scalar_type<S, X>::value && typename enable_if<(has_same_scalar_type<S, X>::value || !is_native_type<S>()) &&
is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_, is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
bool>::type = false> bool>::type = false>
__host__ __device__ constexpr X GetAsType(Idx) const __host__ __device__ constexpr X GetAsType(Idx) const
...@@ -201,7 +201,7 @@ struct StaticTensorTupleOfVectorBuffer ...@@ -201,7 +201,7 @@ struct StaticTensorTupleOfVectorBuffer
// Idx is for S, not X. Idx should be aligned with X // Idx is for S, not X. Idx should be aligned with X
template <typename X, template <typename X,
typename Idx, typename Idx,
typename enable_if<has_same_scalar_type<S, X>::value && typename enable_if<(has_same_scalar_type<S, X>::value || !is_native_type<S>()) &&
is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_, is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
bool>::type = false> bool>::type = false>
__host__ __device__ constexpr void SetAsType(Idx, X x) __host__ __device__ constexpr void SetAsType(Idx, X x)
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp 0 → 100644
View file @ 8b49f207
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_b_scale.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_b_scale.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4_b_scale.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp"
namespace ck {
enum struct BlockGemmPipelineVersion
{
v1, // Naive
v2, // Mem
v3, // Comp
v4, // Comp, double lds buffer
v5, // Comp, double global prefetch register buffer
};
template <BlockGemmPipelineVersion BlkGemmPipelineVer,
BlockGemmPipelineScheduler BlkGemmPipeSche,
index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack>
constexpr auto BlockGemmPipeline_Selector()
{
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
return BlockwiseGemmXdlops_pipeline_v1_b_scale<BlkGemmPipeSche,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
return BlockwiseGemmXdlops_pipeline_v2_b_scale<BlkGemmPipeSche,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
return BlockwiseGemmXdlops_pipeline_v3_b_scale<BlkGemmPipeSche,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
{
return BlockwiseGemmXdlops_pipeline_v4_b_scale<BlkGemmPipeSche,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v5)
{
return BlockwiseGemmXdlops_pipeline_v5<BlkGemmPipeSche,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else
{
std::cerr << "BlockGemmPipeline configuration is not available" << std::endl;
}
}
} // namespace ck
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_b_scale.hpp 0 → 100644
View file @ 8b49f207
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
namespace ck {
// Naive pipeline with lowest resource request per WGP
// GlobalPrefetchStages: 1
// LocalPreFillStages: 1
// LocalPreFetchStages: 0
// LocalSharedMemoryBuffer: 1
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPacks>
struct BlockwiseGemmXdlops_pipeline_v1_b_scale
{
};
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_v1_b_scale<BlockGemmPipelineScheduler::Intrawave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::I0;
using Base::KRepeat;
using Base::xdlops_gemm;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
static constexpr index_t PrefetchStages = 1;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
__host__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
ignore = num_loop;
return TailNumber::Full;
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
// BScale Thread Copy
typename BScaleGridBuffer,
typename BScaleGridDesc,
typename BScaleThreadDesc,
typename BScaleThreadTransfer,
typename BScaleThreadTransferStep>
__device__ void Run(
// ABlockCopy
const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
// BBlockCopy
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
// CThread
CThreadBuffer& c_thread_buf,
// BScaleThreadCopy
const BScaleGridDesc& b_scale_grid_desc,
const BScaleThreadDesc& b_scale_thread_desc,
BScaleThreadTransfer& b_scale_thread_copy,
const BScaleGridBuffer& b_scale_grid_buf,
const BScaleThreadTransferStep& b_scale_thread_copy_step,
// num_loop
index_t num_loop,
index_t num_loop_per_scale) const
{
// assume kperblock = scaleblockk
ignore = num_loop_per_scale;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_scale_thread_desc.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<0>{}));
});
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<1>{}));
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Initialize C
c_thread_buf.Clear();
auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
// -------------------------------------------------------------------------------------------
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
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, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
c_thread_buf_per_scale.Clear();
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(b_scale_thread_buf[n0]);
});
});
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_scale_thread_copy.Run(b_scale_grid_desc,
b_scale_grid_buf,
b_scale_thread_desc,
make_tuple(n0, I0),
b_scale_thread_buf);
b_scale_thread_copy.MoveSrcSliceWindow(
b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
});
b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
b_scale_thread_copy_step.At(Number<1>{}));
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
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, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
c_thread_buf_per_scale.Clear();
static_for<0, KRepeat, 1>{}([&](auto k0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
c_thread_buf(Number<c_offset>{}) +=
c_thread_buf_per_scale[Number<t>{}] *
type_convert<AccDataType>(b_scale_thread_buf[n0]);
});
});
});
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
using Base::b_thread_desc_;
using Base::c_thread_desc_;
};
} // namespace ck
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