Skip to content

GitLab

Commit 43adf1fa authored by Harisankar Sadasivan's avatar Harisankar Sadasivan
Browse files

clang format

parent ab3d3b4a
Show whitespace changes
Inline Side-by-side
Showing with 1154 additions and 1014 deletions
include/ck/host_utility/kernel_launch.hpp
View file @ 43adf1fa
......@@ -11,7 +11,7 @@
#ifndef KERNARG_PRELOAD
template <typename... Args, typename F>
float launch_and_time_kernel(const StreamConfig &stream_config,
float launch_and_time_kernel(const StreamConfig& stream_config,
F kernel,
dim3 grid_dim,
dim3 block_dim,
......@@ -19,7 +19,7 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
Args... args)
{
#if CK_TIME_KERNEL
if (stream_config.time_kernel_)
if(stream_config.time_kernel_)
{
#if DEBUG_LOG
printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
......@@ -49,7 +49,7 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(start, stream_config.stream_id_));
for (int i = 0; i < nrepeat; ++i)
for(int i = 0; i < nrepeat; ++i)
{
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
......@@ -81,7 +81,7 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
#else
template <typename... Args, typename F>
float launch_and_time_kernel(const StreamConfig &stream_config,
float launch_and_time_kernel(const StreamConfig& stream_config,
F kernel,
dim3 grid_dim,
dim3 block_dim,
......@@ -92,7 +92,7 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
// hipGetErrorString(hipMalloc(&args1, sizeof(Args)));
// hip_check_error(hipMemcpy(args1, &args, sizeof(Args), hipMemcpyHostToDevice));
#if CK_TIME_KERNEL
if (stream_config.time_kernel_)
if(stream_config.time_kernel_)
{
#if DEBUG_LOG
printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
......@@ -109,9 +109,9 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
//
// warm up
const int nrepeat = 1000;
for (auto i = 0; i < nrepeat; i++)
hipLaunchKernelGGL(kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_,
args...);
for(auto i = 0; i < nrepeat; i++)
hipLaunchKernelGGL(
kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...);
hip_check_error(hipGetLastError());
#if DEBUG_LOG
......@@ -127,9 +127,9 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
hip_check_error(hipEventRecord(start, stream_config.stream_id_));
for (int i = 0; i < nrepeat; ++i)
hipLaunchKernelGGL(kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_,
args...);
for(int i = 0; i < nrepeat; ++i)
hipLaunchKernelGGL(
kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...);
// hip_check_error(hipGetLastError());
hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
......@@ -140,8 +140,7 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
}
else
{
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(
args...);
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
......@@ -155,7 +154,7 @@ float launch_and_time_kernel(const StreamConfig &stream_config,
}
#endif
template <typename... Args, typename F, typename PreProcessFunc>
float launch_and_time_kernel_with_preprocess(const StreamConfig &stream_config,
float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
PreProcessFunc preprocess,
F kernel,
dim3 grid_dim,
......@@ -164,7 +163,7 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig &stream_config,
Args... args)
{
#if CK_TIME_KERNEL
if (stream_config.time_kernel_)
if(stream_config.time_kernel_)
{
#if DEBUG_LOG
printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
......@@ -195,7 +194,7 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig &stream_config,
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(start, stream_config.stream_id_));
for (int i = 0; i < nrepeat; ++i)
for(int i = 0; i < nrepeat; ++i)
{
preprocess();
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
......
include/ck/tensor_operation/gpu/device/impl/device_tall_and_skinny_gemm_splitk.hpp
View file @ 43adf1fa
......@@ -16,14 +16,11 @@
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck
{
namespace tensor_operation
{
namespace device
{
namespace ck {
namespace tensor_operation {
namespace device {
template <
template <
typename ADataType,
typename BDataType,
typename CDataType,
......@@ -61,7 +58,7 @@ namespace ck
is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
bool> = false>
struct deviceTsmmDl : public DeviceTsmm<ALayout,
struct deviceTsmmDl : public DeviceTsmm<ALayout,
BLayout,
CLayout,
ADataType,
......@@ -71,7 +68,7 @@ namespace ck
BElementwiseOperation,
CElementwiseOperation>
{
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
......@@ -116,7 +113,7 @@ namespace ck
struct Invoker : public BaseInvoker
{
float Run(const Argument &karg, const StreamConfig &stream_config = StreamConfig{})
float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
{
const index_t grid_size = GridwiseTsmm::CalculateGridSize(karg.M, karg.N, karg.k_batch);
......@@ -130,12 +127,12 @@ namespace ck
float ave_time = 0;
if (karg.k_batch > 1)
if(karg.k_batch > 1)
hipGetErrorString(hipMemset(karg.p_c_grid, 0, karg.M * karg.N * sizeof(CDataType)));
if (has_main_k_block_loop && has_double_tail_k_block_loop)
if(has_main_k_block_loop && has_double_tail_k_block_loop)
{
if (karg.k_batch == 1)
if(karg.k_batch == 1)
{
const auto kernel = kernel_tsmm_dl_v1r3<GridwiseTsmm,
......@@ -146,9 +143,22 @@ namespace ck
true,
true,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
else
{
......@@ -160,15 +170,28 @@ namespace ck
true,
true,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
}
else if (has_main_k_block_loop && !has_double_tail_k_block_loop)
else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
{
if (karg.k_batch == 1)
if(karg.k_batch == 1)
{
const auto kernel = kernel_tsmm_dl_v1r3<GridwiseTsmm,
ADataType,
......@@ -178,9 +201,22 @@ namespace ck
true,
false,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
else
{
......@@ -192,14 +228,27 @@ namespace ck
true,
false,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
}
else if (!has_main_k_block_loop && has_double_tail_k_block_loop)
else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
{
if (karg.k_batch == 1)
if(karg.k_batch == 1)
{
const auto kernel = kernel_tsmm_dl_v1r3<GridwiseTsmm,
ADataType,
......@@ -209,9 +258,22 @@ namespace ck
false,
true,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
else
{
......@@ -223,14 +285,27 @@ namespace ck
false,
true,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
}
else
{
if (karg.k_batch == 1)
if(karg.k_batch == 1)
{
const auto kernel = kernel_tsmm_dl_v1r3<GridwiseTsmm,
ADataType,
......@@ -240,9 +315,22 @@ namespace ck
false,
false,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
else
{
......@@ -254,20 +342,32 @@ namespace ck
false,
false,
DefaultBlock2CTileMap>; // //
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(grid_size), dim3(BlockSize), 0, karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, (karg.M), (karg.N), (karg.K),
(karg.K0), (karg.k_batch), karg.MPadded, karg.NPadded, b2c_map);
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
(karg.M),
(karg.N),
(karg.K),
(karg.K0),
(karg.k_batch),
karg.MPadded,
karg.NPadded,
b2c_map);
}
}
return ave_time;
}
// polymorphic
float
Run(const BaseArgument *p_arg,
const StreamConfig &stream_config = StreamConfig{}) override
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument *>(p_arg), stream_config);
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
......@@ -277,12 +377,13 @@ namespace ck
return true;
}
// //
static bool IsSupportedArgument(const Argument &arg)
static bool IsSupportedArgument(const Argument& arg)
{
if (ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102" || ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942")
ck::get_device_name() == "gfx1102" || ck::get_device_name() == "gfx940" ||
ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942")
{
return GridwiseTsmm::CheckValidity(arg);
}
......@@ -293,14 +394,14 @@ namespace ck
}
// //
// polymorphic
bool IsSupportedArgument(const BaseArgument *p_arg) override
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument *>(p_arg));
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType *p_a,
const BDataType *p_b,
CDataType *p_c,
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
index_t M,
index_t N,
index_t K,
......@@ -331,9 +432,9 @@ namespace ck
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void *p_a,
const void *p_b,
void *p_c,
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
index_t M,
index_t N,
index_t K,
......@@ -346,9 +447,9 @@ namespace ck
ck::index_t KBatch = 1) override // //
{
return std::make_unique<Argument>(static_cast<const ADataType *>(p_a),
static_cast<const BDataType *>(p_b),
static_cast<CDataType *>(p_c),
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
M,
N,
K,
......@@ -389,8 +490,8 @@ namespace ck
return str.str();
}
};
};
} // namespace device
} // namespace tensor_operation
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_tall_and_skinny_gemm_splitk.hpp
View file @ 43adf1fa
......@@ -16,10 +16,9 @@
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck
{
namespace ck {
template <typename GridwiseTsmm,
template <typename GridwiseTsmm,
typename FloatAB,
typename FloatC,
typename BLayout,
......@@ -27,35 +26,68 @@ namespace ck
bool HasMainKBlockLoop,
bool HasDoubleTailKBlockLoop,
typename Block2CTileMap>
__global__ void
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_tsmm_dl_v1r3(
const FloatAB *p_a_grid, const FloatAB *p_b_grid, FloatC *p_c_grid, index_t M, index_t N, index_t K,
index_t K0, index_t k_batch, index_t MPadded, index_t NPadded, const Block2CTileMap block_2_ctile_map) //: in __global__ functions, struct is
const FloatAB* p_a_grid,
const FloatAB* p_b_grid,
FloatC* p_c_grid,
index_t M,
index_t N,
index_t K,
index_t K0,
index_t k_batch,
index_t MPadded,
index_t NPadded,
const Block2CTileMap block_2_ctile_map) //: in __global__ functions, struct is
// better for reduced load overhead
{
{
// strides depend on B's layout
if constexpr (is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
GridwiseTsmm::template Run<HasMainKBlockLoop,
HasDoubleTailKBlockLoop,
GridwiseTsmm,
CGlobalMemoryDataOperation>(p_a_grid, p_b_grid, p_c_grid, M, N, K,
K0, k_batch, K, N, N, MPadded, NPadded, block_2_ctile_map);
CGlobalMemoryDataOperation>(p_a_grid,
p_b_grid,
p_c_grid,
M,
N,
K,
K0,
k_batch,
K,
N,
N,
MPadded,
NPadded,
block_2_ctile_map);
}
else
{
GridwiseTsmm::template Run<HasMainKBlockLoop,
HasDoubleTailKBlockLoop,
GridwiseTsmm,
CGlobalMemoryDataOperation>(p_a_grid, p_b_grid, p_c_grid, M, N, K,
K0, k_batch, K, K, N, MPadded, NPadded, block_2_ctile_map);
}
}
template <index_t BlockSize,
CGlobalMemoryDataOperation>(p_a_grid,
p_b_grid,
p_c_grid,
M,
N,
K,
K0,
k_batch,
K,
K,
N,
MPadded,
NPadded,
block_2_ctile_map);
}
}
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
......@@ -83,8 +115,8 @@ namespace ck
typename CThreadTransferSrcDstAccessOrder,
index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector>
struct GridwiseTsmmDl_km_kn_mn
{
struct GridwiseTsmmDl_km_kn_mn
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
......@@ -96,9 +128,9 @@ namespace ck
// Argument
struct Argument : public tensor_operation::device::BaseArgument //
{
Argument(const FloatAB *p_a_grid_,
const FloatAB *p_b_grid_,
FloatC *p_c_grid_,
Argument(const FloatAB* p_a_grid_,
const FloatAB* p_b_grid_,
FloatC* p_c_grid_,
index_t M_,
index_t N_,
index_t K_,
......@@ -128,9 +160,9 @@ namespace ck
}
// private:
const FloatAB *p_a_grid;
const FloatAB *p_b_grid;
FloatC *p_c_grid;
const FloatAB* p_a_grid;
const FloatAB* p_b_grid;
FloatC* p_c_grid;
index_t M, N, K;
index_t StrideA, StrideB, StrideC;
......@@ -214,19 +246,18 @@ namespace ck
index_t M, index_t MPad, index_t K, index_t StrideA, index_t KBatch, index_t K0)
{
const auto a_grid_desc_m_k = [&]()
{
if constexpr (is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr (is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
if constexpr (GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
return transform_tensor_descriptor(
......@@ -255,19 +286,18 @@ namespace ck
index_t K, index_t NPad, index_t N, index_t StrideB, index_t KBatch, index_t K0)
{
const auto b_grid_desc_k_n = [&]()
{
if constexpr (is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr (is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
if constexpr (GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
return transform_tensor_descriptor(
......@@ -290,19 +320,18 @@ namespace ck
__host__ __device__ static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]()
{
if constexpr (is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
const auto c_grid_desc_m_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr (is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
if constexpr (GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
......@@ -335,7 +364,7 @@ namespace ck
using BGridDesc_Kbatch_K0_N_K1 = decltype(MakeBGridDescriptor_KBatch_K0_N_K1(1, 1, 1, 1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
__host__ __device__ static constexpr bool CheckValidity(const Argument &karg)
__host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
{
// const auto MPadded = CalculateMPadded(karg.M);
......@@ -361,7 +390,7 @@ namespace ck
// KBatch, K0, M, K1 -> KBatch, K0, M0, M1 (MPerBlock), K1
__host__ __device__ static constexpr auto MakeAGridDescriptor_Kbatch_K0_M0_M1_K1(
const AGridDesc_Kbatch_K0_M_K1 &a_grid_desc_kbatch_k0_m_k1)
const AGridDesc_Kbatch_K0_M_K1& a_grid_desc_kbatch_k0_m_k1)
{
const auto KBatch = a_grid_desc_kbatch_k0_m_k1.GetLength(I0);
const auto K0 = a_grid_desc_kbatch_k0_m_k1.GetLength(I1);
......@@ -383,7 +412,7 @@ namespace ck
}
__host__ __device__ static constexpr auto MakeBGridDescriptor_Kbatch_K0_N0_N1_K1(
const BGridDesc_Kbatch_K0_N_K1 &b_grid_desc_kbatch_k0_n_k1)
const BGridDesc_Kbatch_K0_N_K1& b_grid_desc_kbatch_k0_n_k1)
{
const auto KBatch = b_grid_desc_kbatch_k0_n_k1.GetLength(I0);
const auto K0 = b_grid_desc_kbatch_k0_n_k1.GetLength(I1);
......@@ -405,7 +434,7 @@ namespace ck
}
__host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N &c_grid_desc_m_n)
MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N& c_grid_desc_m_n)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
......@@ -451,8 +480,20 @@ namespace ck
bool HasDoubleTailKBlockLoop,
typename GridwiseTsmm,
InMemoryDataOperationEnum CGlobalMemoryDataOperation>
__device__ static void Run(const FloatAB *p_a_grid, const FloatAB *p_b_grid, FloatC *p_c_grid, index_t M, index_t N, index_t K,
index_t K0, index_t k_batch, index_t StrideA, index_t StrideB, index_t StrideC, index_t MPadded, index_t NPadded, const Block2CTileMap &block_2_ctile_map)
__device__ static void Run(const FloatAB* p_a_grid,
const FloatAB* p_b_grid,
FloatC* p_c_grid,
index_t M,
index_t N,
index_t K,
index_t K0,
index_t k_batch,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t MPadded,
index_t NPadded,
const Block2CTileMap& block_2_ctile_map)
{
constexpr index_t shared_block_size =
......@@ -464,8 +505,7 @@ namespace ck
M, MPadded, K, StrideA, k_batch, K0); //
const auto b_grid_desc_kbatch_k0_n_k1 = GridwiseTsmm::MakeBGridDescriptor_KBatch_K0_N_K1(
K, NPadded, N, StrideB, k_batch, K0); //
const auto c_grid_desc_m_n =
GridwiseTsmm::MakeCGridDescriptor_M_N(M, N, StrideC);
const auto c_grid_desc_m_n = GridwiseTsmm::MakeCGridDescriptor_M_N(M, N, StrideC);
const auto a_grid_desc_kbatch_k0_m0_m1_k1 =
GridwiseTsmm::MakeAGridDescriptor_Kbatch_K0_M0_M1_K1(a_grid_desc_kbatch_k0_m_k1); //
......@@ -482,15 +522,15 @@ namespace ck
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
const auto c_m0_n0_block_cluster_idx = block_2_ctile_map.convert_1D_block_idx_to_3D_tuple(
get_block_1d_id(), N, k_batch);
const auto c_m0_n0_block_cluster_idx =
block_2_ctile_map.convert_1D_block_idx_to_3D_tuple(get_block_1d_id(), N, k_batch);
// HACK: this force index data into SGPR
const index_t im0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I0]);
const index_t in0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I1]);
const index_t kbatch_id = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I2]);
if (!block_2_ctile_map.ValidCTileIndex(
if(!block_2_ctile_map.ValidCTileIndex(
make_tuple(im0, in0),
make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
......@@ -593,7 +633,7 @@ namespace ck
constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
FloatAB *p_a_block_double = p_shared_block;
FloatAB* p_a_block_double = p_shared_block;
auto b_thread_odd_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
b_k0_n_k1_thread_desc.GetElementSpaceSize());
......@@ -632,7 +672,7 @@ namespace ck
b_thread_even_buf);
}
if constexpr (HasMainKBlockLoop)
if constexpr(HasMainKBlockLoop)
{
// const auto K0 = a_grid_desc_kbatch_k0_m0_m1_k1.GetLength(I1);
......@@ -691,11 +731,11 @@ namespace ck
a_blockwise_copy.RunWrite(a_block_desc_copy_kbatch_k0_m0_m1_k1, a_block_even_buf);
k_block_data_begin += 2 * K0PerBlock;
} while (k_block_data_begin < K0 - 2 * K0PerBlock);
} while(k_block_data_begin < K0 - 2 * K0PerBlock);
}
// LDS double buffer: tail
if constexpr (HasDoubleTailKBlockLoop) // if has 2 iteration left
if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
{
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_kbatch_k0_m0_m1_k1,
a_block_slice_copy_step);
......@@ -780,5 +820,5 @@ namespace ck
c_grid_buf);
}
}
};
};
} // 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