mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-04 13:41:24 +00:00
support flatmm scaling
This commit is contained in:
Feng Shijie
@@ -23,9 +23,12 @@ template <typename FlatmmConfig,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename ScaleM,
|
||||
typename ScaleN,
|
||||
bool persistent,
|
||||
typename CDEElementWise>
|
||||
float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_config& s)
|
||||
float flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
|
||||
const ck_tile::stream_config& s)
|
||||
{
|
||||
using CodegenFlatmmShape = ck_tile::TileGemmShape<
|
||||
ck_tile::sequence<FlatmmConfig::M_Tile, FlatmmConfig::N_Tile, FlatmmConfig::K_Tile>,
|
||||
@@ -81,13 +84,13 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
|
||||
constexpr auto memory_operation = memory_operation_.value;
|
||||
|
||||
using CodegenPipelineProblem = ck_tile::FlatmmPipelineProblem<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CodegenFlatmmShape,
|
||||
CodegenGemmTraits,
|
||||
scheduler,
|
||||
has_hot_loop_v,
|
||||
tail_number_v>;
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CodegenFlatmmShape,
|
||||
CodegenGemmTraits,
|
||||
scheduler,
|
||||
has_hot_loop_v,
|
||||
tail_number_v>;
|
||||
|
||||
using CodegenFlatmmPipeline =
|
||||
ck_tile::FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
|
||||
@@ -217,6 +220,7 @@ int run_flatmm_example(int argc, char* argv[])
|
||||
std::string data_type = arg_parser.get_str("prec");
|
||||
std::string a_layout = arg_parser.get_str("a_layout");
|
||||
std::string b_layout = arg_parser.get_str("b_layout");
|
||||
int scale_opt = arg_parser.get_int("scale");
|
||||
if(a_layout == "R" && b_layout == "C")
|
||||
{
|
||||
if(data_type == "fp16")
|
||||
@@ -231,13 +235,29 @@ int run_flatmm_example(int argc, char* argv[])
|
||||
}
|
||||
else if(data_type == "fp8")
|
||||
{
|
||||
run_flatmm_example_with_layouts<ck_tile::fp8_t, FlatmmConfig<ck_tile::fp8_t>>(
|
||||
argc, argv, Row{}, Col{}, Row{});
|
||||
if(scale_opt == 0)
|
||||
{
|
||||
run_flatmm_example_with_layouts<ck_tile::fp8_t, FlatmmConfig<ck_tile::fp8_t>>(
|
||||
argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
run_flatmm_example_with_layouts<ck_tile::fp8_t, FlatmmConfig<ck_tile::fp8_t>, 1, 1>(
|
||||
argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
}
|
||||
else if(data_type == "bf8")
|
||||
{
|
||||
run_flatmm_example_with_layouts<ck_tile::bf8_t, FlatmmConfig<ck_tile::bf8_t>>(
|
||||
argc, argv, Row{}, Col{}, Row{});
|
||||
if(scale_opt == 0)
|
||||
{
|
||||
run_flatmm_example_with_layouts<ck_tile::bf8_t, FlatmmConfig<ck_tile::bf8_t>>(
|
||||
argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
else
|
||||
{
|
||||
run_flatmm_example_with_layouts<ck_tile::bf8_t, FlatmmConfig<ck_tile::bf8_t>, 1, 1>(
|
||||
argc, argv, Row{}, Col{}, Row{});
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
|
||||
@@ -83,10 +83,10 @@ struct FlatmmConfig16
|
||||
template <typename DataType>
|
||||
struct FlatmmConfig16_950 : public FlatmmConfig16<DataType>
|
||||
{
|
||||
static constexpr ck_tile::index_t N_Tile = 256;
|
||||
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(DataType);
|
||||
static constexpr ck_tile::index_t N_Tile = 256;
|
||||
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(DataType);
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(DataType) == 2 ? 32 : 128;
|
||||
static constexpr int kBlockPerCu = 1;
|
||||
static constexpr int kBlockPerCu = 1;
|
||||
};
|
||||
|
||||
template <typename ADataType>
|
||||
@@ -167,120 +167,6 @@ struct is_8bit_type
|
||||
{
|
||||
};
|
||||
|
||||
template <typename DataType>
|
||||
struct GemmConfig
|
||||
{
|
||||
#if defined(USING_MFMA_16x16x128_F8) //MI350 FP8 16X16
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 256;
|
||||
static constexpr ck_tile::index_t K_Tile = 256;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 128;
|
||||
#elif defined(USING_MFMA_32x32x64_F8) //MI350 FP8 32X32 (need tune)
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 128;
|
||||
static constexpr ck_tile::index_t K_Tile = 128;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 64;
|
||||
#elif defined(USING_MFMA_16x16x32_F16) //MI350 FP16 16X16 (need tune)
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 128;
|
||||
static constexpr ck_tile::index_t K_Tile = 128;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 32;
|
||||
#elif defined(USING_MFMA_32x32x16_F16) //MI350 FP16 32X32 (need tune)
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 128;
|
||||
static constexpr ck_tile::index_t K_Tile = 128;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
#elif defined(USING_MFMA_16x16x32_F8) //MI300 FP8 16X16
|
||||
static constexpr ck_tile::index_t M_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Tile = 64;
|
||||
static constexpr ck_tile::index_t K_Tile = 256;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 64;
|
||||
#elif defined(USING_MFMA_32x32x16_F8) //MI300 FP8 32X32 (need tune)
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 256;
|
||||
static constexpr ck_tile::index_t K_Tile = 128;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 8;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 32;
|
||||
#elif defined(USING_MFMA_16x16x16_F16) //MI300 FP16 16X16 (need tune)
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 128;
|
||||
static constexpr ck_tile::index_t K_Tile = 128;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 32;
|
||||
#elif defined(USING_MFMA_32x32x8_F16) //MI300 FP16 32X32 (need tune)
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 128;
|
||||
static constexpr ck_tile::index_t K_Tile = 128;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 32;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 16;
|
||||
#else
|
||||
static constexpr ck_tile::index_t M_Tile = 128;
|
||||
static constexpr ck_tile::index_t N_Tile = 256;
|
||||
static constexpr ck_tile::index_t K_Tile = 256;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp = 1;
|
||||
static constexpr ck_tile::index_t N_Warp = 4;
|
||||
static constexpr ck_tile::index_t K_Warp = 1;
|
||||
|
||||
static constexpr ck_tile::index_t M_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t N_Warp_Tile = 16;
|
||||
static constexpr ck_tile::index_t K_Warp_Tile = 128;
|
||||
#endif
|
||||
};
|
||||
|
||||
auto create_args(int argc, char* argv[])
|
||||
{
|
||||
ck_tile::ArgParser arg_parser;
|
||||
@@ -301,6 +187,7 @@ auto create_args(int argc, char* argv[])
|
||||
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
|
||||
.insert("split_k", "1", "splitK value")
|
||||
.insert("init", "0", "0:random, 1:linear, 2:constant(1)")
|
||||
.insert("scale", "0", "0:without scale, 1:per-token/channel scale, only for fp8/bf8")
|
||||
.insert("warp_tile",
|
||||
"0",
|
||||
"0: 16x16, 1: 32x32, 2: 16x16x128 (950 only), 3: 32x32x64 (950 only)");
|
||||
|
||||
@@ -18,7 +18,7 @@ constexpr const char* DataTypeToString()
|
||||
{
|
||||
return "bf8";
|
||||
}
|
||||
else if constexpr(std::is_same_v<T, ck_tile::bf16_t>)
|
||||
else if constexpr(std::is_same_v<T, ck_tile::bf16_t>)
|
||||
{
|
||||
return "bf16";
|
||||
}
|
||||
@@ -83,9 +83,12 @@ template <typename FlatmmConfig,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename ScaleM,
|
||||
typename ScaleN,
|
||||
bool persistent,
|
||||
typename CDEElementWise>
|
||||
float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_config& s);
|
||||
float flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
|
||||
const ck_tile::stream_config& s);
|
||||
|
||||
template <typename FlatmmConfig,
|
||||
typename ADataType,
|
||||
@@ -97,6 +100,8 @@ template <typename FlatmmConfig,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename CLayout,
|
||||
typename ScaleM,
|
||||
typename ScaleN,
|
||||
typename CDEElementWise = ck_tile::element_wise::PassThrough>
|
||||
float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
|
||||
ck_tile::DeviceMem& b_shuffle_dev_buf,
|
||||
@@ -108,21 +113,25 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
|
||||
ck_tile::index_t stride_B,
|
||||
ck_tile::index_t stride_C,
|
||||
ck_tile::index_t kbatch,
|
||||
ScaleM scale_m,
|
||||
ScaleN scale_n,
|
||||
int n_warmup,
|
||||
int n_repeat)
|
||||
{
|
||||
ck_tile::FlatmmHostArgs<> args = {a_dev_buf.GetDeviceBuffer(),
|
||||
b_shuffle_dev_buf.GetDeviceBuffer(),
|
||||
{},
|
||||
c_dev_buf.GetDeviceBuffer(),
|
||||
kbatch,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
{},
|
||||
stride_C};
|
||||
ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN> args = {a_dev_buf.GetDeviceBuffer(),
|
||||
b_shuffle_dev_buf.GetDeviceBuffer(),
|
||||
{},
|
||||
c_dev_buf.GetDeviceBuffer(),
|
||||
kbatch,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
{},
|
||||
stride_C,
|
||||
scale_m,
|
||||
scale_n};
|
||||
|
||||
float ave_time = flatmm_calc<FlatmmConfig,
|
||||
ADataType,
|
||||
@@ -134,6 +143,8 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
CLayout,
|
||||
ScaleM,
|
||||
ScaleN,
|
||||
false,
|
||||
CDEElementWise>(
|
||||
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
|
||||
@@ -154,6 +165,8 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
|
||||
|
||||
template <typename PrecType,
|
||||
typename FlatmmConfig,
|
||||
int ScaleGranularityM = -1,
|
||||
int ScaleGranularityN = -1,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
@@ -197,21 +210,30 @@ int run_flatmm_example_with_layouts(int argc,
|
||||
ck_tile::HostTensor<CDataType> c_rslt_host(
|
||||
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
||||
|
||||
ck_tile::HostTensor<AccDataType> per_token_scale(ck_tile::HostTensorDescriptor({M}, {1}));
|
||||
ck_tile::HostTensor<AccDataType> per_channel_scale(ck_tile::HostTensorDescriptor({N}, {1}));
|
||||
|
||||
// TODO: add different init types
|
||||
if(init_method == 0)
|
||||
{
|
||||
ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
|
||||
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
|
||||
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_token_scale);
|
||||
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_channel_scale);
|
||||
}
|
||||
else if(init_method == 1)
|
||||
{
|
||||
ck_tile::FillMonotonicSeq<ADataType>{}(a_host);
|
||||
ck_tile::FillMonotonicSeq<BDataType>{}(b_origin_host);
|
||||
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_token_scale);
|
||||
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_channel_scale);
|
||||
}
|
||||
else if(init_method == 2)
|
||||
{
|
||||
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_host);
|
||||
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
|
||||
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_token_scale);
|
||||
ck_tile::FillUniformDistribution<AccDataType>{1.f, 1.f}(per_channel_scale);
|
||||
}
|
||||
else
|
||||
{
|
||||
@@ -222,14 +244,25 @@ int run_flatmm_example_with_layouts(int argc,
|
||||
ck_tile::DeviceMem a_dev_buf(a_host.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem c_dev_buf(c_rslt_host.get_element_space_size_in_bytes());
|
||||
|
||||
ck_tile::DeviceMem per_token_scale_dev_buf(per_token_scale.get_element_space_size_in_bytes());
|
||||
ck_tile::DeviceMem per_channel_scale_dev_buf(
|
||||
per_channel_scale.get_element_space_size_in_bytes());
|
||||
|
||||
a_dev_buf.ToDevice(a_host.data());
|
||||
c_rslt_host.SetZero();
|
||||
per_token_scale_dev_buf.ToDevice(per_token_scale.data());
|
||||
per_channel_scale_dev_buf.ToDevice(per_channel_scale.data());
|
||||
|
||||
// do pre-shuffle
|
||||
ck_tile::HostTensor<BDataType> b_shuffle_host = shuffle_b<FlatmmConfig>(b_origin_host);
|
||||
ck_tile::DeviceMem b_shuffle_dev_buf(b_shuffle_host.get_element_space_size_in_bytes());
|
||||
b_shuffle_dev_buf.ToDevice(b_shuffle_host.data());
|
||||
|
||||
auto per_token_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityM>{
|
||||
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer())};
|
||||
auto per_channel_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityN>{
|
||||
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
|
||||
|
||||
invoke_flatmm<FlatmmConfig,
|
||||
ADataType,
|
||||
BDataType,
|
||||
@@ -239,18 +272,22 @@ int run_flatmm_example_with_layouts(int argc,
|
||||
ALayout,
|
||||
BLayout,
|
||||
ck_tile::tuple<>,
|
||||
CLayout>(a_dev_buf,
|
||||
b_shuffle_dev_buf,
|
||||
c_dev_buf,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
stride_C,
|
||||
kbatch,
|
||||
n_warmup,
|
||||
n_repeat);
|
||||
CLayout,
|
||||
decltype(per_token_scale_dev_ptr),
|
||||
decltype(per_channel_scale_dev_ptr)>(a_dev_buf,
|
||||
b_shuffle_dev_buf,
|
||||
c_dev_buf,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
stride_C,
|
||||
kbatch,
|
||||
per_token_scale_dev_ptr,
|
||||
per_channel_scale_dev_ptr,
|
||||
n_warmup,
|
||||
n_repeat);
|
||||
|
||||
c_dev_buf.FromDevice(c_rslt_host.data());
|
||||
bool pass = true;
|
||||
@@ -263,6 +300,8 @@ int run_flatmm_example_with_layouts(int argc,
|
||||
|
||||
if(arg_parser.get_int("v") == 1)
|
||||
{
|
||||
assert(ScaleGranularityM == -1 && ScaleGranularityN == -1 &&
|
||||
"ScaleAB is not supported for CPU verification!");
|
||||
ck_tile::HostTensor<CDataType> c_ref_host(
|
||||
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
||||
c_ref_host.SetZero();
|
||||
@@ -310,13 +349,41 @@ int run_flatmm_example_with_layouts(int argc,
|
||||
N * K * sizeof(BDataType),
|
||||
hipMemcpyHostToDevice));
|
||||
|
||||
ck_tile::reference_gemm_gpu<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
|
||||
if constexpr(ScaleGranularityM == -1 && ScaleGranularityN == -1)
|
||||
{
|
||||
ck_tile::reference_gemm_gpu<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(
|
||||
d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
|
||||
}
|
||||
else
|
||||
{
|
||||
ck_tile::reference_blockwise_gemm_gpu<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CDataType,
|
||||
ALayout,
|
||||
BLayout,
|
||||
CLayout>(
|
||||
d_A,
|
||||
d_B,
|
||||
d_C,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
stride_A,
|
||||
stride_B,
|
||||
stride_C,
|
||||
ScaleGranularityM,
|
||||
ScaleGranularityN,
|
||||
K,
|
||||
static_cast<float*>(per_token_scale_dev_buf.GetDeviceBuffer()),
|
||||
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer()));
|
||||
}
|
||||
|
||||
ck_tile::hip_check_error(hipMemcpy(c_gpu_ref_dev_buf.GetDeviceBuffer(),
|
||||
d_C,
|
||||
|
||||
Reference in New Issue
Block a user