ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Conv:TF32: add more instances - 2 (#2879)

Browse Source 
* add instances of device_grouped_conv_fwd_xdl_f32_comp_instances
* add instances of device_grouped_conv_fwd_xdl_f32_tf32_mem_instances
* add instances of device_grouped_conv_fwd_xdl_large_tensor_f32_tf32_instances
* tf32:conv:add instances for base class DeviceConvFwd
* tf32:conv:add instances for base class DeviceGroupedConvBwdDataMultipleD
* tf32:conv:add instances for base class DeviceGroupedConvBwdWeight
* add tf32 in profiler
* remove gnhwc/ngchw/ngcdhw instances
* remove non-ndhwgc/nhwgc/nhwc instances
* add check in IsSupportedArgument()
This commit is contained in:
yinglu
2025-10-10 15:28:17 +08:00
committed by GitHub
parent ad7a215aba
commit fada1a3cae
56 changed files with 2119 additions and 152 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
View File

@@ -29,7 +29,8 @@ template <ck::index_t NDimSpatial,
typename InLayout,
typename OutDataType,
typename WeiDataType,
typename InDataType>
typename InDataType,
typename ComputeDataType = InDataType>
bool profile_grouped_conv_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
@@ -96,7 +97,11 @@ bool profile_grouped_conv_bwd_data_impl(int do_verification,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
OutElementOp,
0,
0,
0,
ComputeDataType>();
auto ref_invoker = ref_conv.MakeInvoker();
@@ -171,9 +176,13 @@ bool profile_grouped_conv_bwd_data_impl(int do_verification,
{
in_device_buf.FromDevice(in_device.mData.data());
using ComputeType = std::conditional_t<sizeof(OutDataType) < sizeof(WeiDataType),
OutDataType,
WeiDataType>;
using ComputeType_ = std::conditional_t<sizeof(OutDataType) < sizeof(WeiDataType),
OutDataType,
WeiDataType>;
using ComputeType =
std::conditional_t<sizeof(ComputeType_) < sizeof(ComputeDataType),
ComputeType_,
ComputeDataType>;
using AccDataType =
std::conditional_t<std::is_same_v<ComputeType, int8_t>, int32_t, float>;
const index_t num_accums = conv_param.K_;
@@ -222,18 +231,21 @@ bool profile_grouped_conv_bwd_data_impl(int do_verification,
};
// do GEMM
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
OutElementOp,
WeiElementOp,
InElementOp>;
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
OutElementOp,
WeiElementOp,
InElementOp,
ComputeDataType,
ComputeDataType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<

120
profiler/src/profile_grouped_conv_bwd_data.cpp
View File

@@ -21,9 +21,10 @@ enum struct ConvLayout
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
F32_F32_F32_TF32, // 3
};
#define OP_NAME "grouped_conv_bwd_data"
@@ -37,6 +38,7 @@ static void print_helper_msg()
<< "arg2: data type (0: Output fp32, Weight fp32, Input fp32\n"
<< " 1: Output fp16, Weight fp16, Input fp16\n"
<< " 2: Output bf16, Weight bf16, Input bf16\n"
<< " 3: Output fp32, Weight fp32, Input fp32, Compute tf32)\n"
<< "arg3: tensor layout (0: Output[G, N, Ho, Wo, C], Weight[G, K, Y, X, C], Input[G, N, Hi, Wi, K]\n"
<< " 1: Output[N, Ho, Wo, G, C], Weight[G, K, Y, X, C], Input[N, Hi, Wi, G, K])\n"
<< " 2: Output[N, G, C, Ho, Wo], Weight[G, K, Y, X, C], Input[N, G, K, Hi, Wi])\n"
@@ -82,6 +84,9 @@ int profile_grouped_conv_bwd_data(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#if defined(__gfx942__)
using TF32 = ck::tf32_t;
#endif
using namespace ck::tensor_layout::convolution;
@@ -94,16 +99,18 @@ int profile_grouped_conv_bwd_data(int argc, char* argv[])
auto in_layout,
auto wei_type,
auto out_type,
auto in_type) {
auto in_type,
auto compute_type) {
constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
using OutLayout = decltype(out_layout);
using WeiLayout = decltype(wei_layout);
using InLayout = decltype(in_layout);
using OutDataType = decltype(out_type);
using WeiDataType = decltype(wei_type);
using InDataType = decltype(in_type);
using OutDataType = decltype(out_type);
using WeiDataType = decltype(wei_type);
using InDataType = decltype(in_type);
using ComputeDataType = decltype(compute_type);
bool pass = ck::profiler::profile_grouped_conv_bwd_data_impl<NDimSpatial,
OutLayout,
@@ -111,7 +118,8 @@ int profile_grouped_conv_bwd_data(int argc, char* argv[])
InLayout,
OutDataType,
WeiDataType,
InDataType>(
InDataType,
ComputeDataType>(
do_verification, init_method, do_log, time_kernel, params, split_k);
return pass ? 0 : 1;
@@ -123,60 +131,84 @@ int profile_grouped_conv_bwd_data(int argc, char* argv[])
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F32{}, F32{}, F32{});
return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F16{}, F16{}, F16{});
return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, BF16{}, BF16{}, BF16{});
return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
else if(layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F32{}, F32{}, F32{});
return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F16{}, F16{}, F16{});
return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, BF16{}, BF16{}, BF16{});
return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
else if(layout == ConvLayout::NGCHW_GKYXC_NGKHW)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NGKHW{}, GKYXC{}, NGCHW{}, F32{}, F32{}, F32{});
return profile(I2, NGKHW{}, GKYXC{}, NGCHW{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NGKHW{}, GKYXC{}, NGCHW{}, F16{}, F16{}, F16{});
return profile(I2, NGKHW{}, GKYXC{}, NGCHW{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I2, NGKHW{}, GKYXC{}, NGCHW{}, BF16{}, BF16{}, BF16{});
return profile(I2, NGKHW{}, GKYXC{}, NGCHW{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NGKHW{}, GKYXC{}, NGCHW{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
else if(layout == ConvLayout::NGCHW_GKCYX_NGKHW)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NGKHW{}, GKCYX{}, NGCHW{}, F32{}, F32{}, F32{});
return profile(I2, NGKHW{}, GKCYX{}, NGCHW{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NGKHW{}, GKCYX{}, NGCHW{}, F16{}, F16{}, F16{});
return profile(I2, NGKHW{}, GKCYX{}, NGCHW{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I2, NGKHW{}, GKCYX{}, NGCHW{}, BF16{}, BF16{}, BF16{});
return profile(I2, NGKHW{}, GKCYX{}, NGCHW{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NGKHW{}, GKCYX{}, NGCHW{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
}
@@ -186,60 +218,84 @@ int profile_grouped_conv_bwd_data(int argc, char* argv[])
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, F32{}, F32{}, F32{});
return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, F16{}, F16{}, F16{});
return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, BF16{}, BF16{}, BF16{});
return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
else if(layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, F32{}, F32{}, F32{});
return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, F16{}, F16{}, F16{});
return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, BF16{}, BF16{}, BF16{});
return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
else if(layout == ConvLayout::NGCHW_GKYXC_NGKHW)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, NGKDHW{}, GKZYXC{}, NGCDHW{}, F32{}, F32{}, F32{});
return profile(I3, NGKDHW{}, GKZYXC{}, NGCDHW{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, NGKDHW{}, GKZYXC{}, NGCDHW{}, F16{}, F16{}, F16{});
return profile(I3, NGKDHW{}, GKZYXC{}, NGCDHW{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I3, NGKDHW{}, GKZYXC{}, NGCDHW{}, BF16{}, BF16{}, BF16{});
return profile(I3, NGKDHW{}, GKZYXC{}, NGCDHW{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, NGKDHW{}, GKZYXC{}, NGCDHW{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
else if(layout == ConvLayout::NGCHW_GKYXC_NGKHW)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, NGKDHW{}, GKCZYX{}, NGCDHW{}, F32{}, F32{}, F32{});
return profile(I3, NGKDHW{}, GKCZYX{}, NGCDHW{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, NGKDHW{}, GKCZYX{}, NGCDHW{}, F16{}, F16{}, F16{});
return profile(I3, NGKDHW{}, GKCZYX{}, NGCDHW{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16)
{
return profile(I3, NGKDHW{}, GKCZYX{}, NGCDHW{}, BF16{}, BF16{}, BF16{});
return profile(I3, NGKDHW{}, GKCZYX{}, NGCDHW{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, NGKDHW{}, GKCZYX{}, NGCDHW{}, F32{}, F32{}, F32{}, TF32{});
#endif
}
}
}

49
profiler/src/profile_grouped_conv_bwd_weight.cpp
View File

@@ -28,6 +28,7 @@ enum struct ConvDataType
F16_F16_F16_BF8_F8, // 3
I8_I8_I8, // 4
BF16_BF16_BF16, // 5
F32_F32_F32_TF32, // 6
};
#define OP_NAME "grouped_conv_bwd_weight"
@@ -41,7 +42,8 @@ static void print_helper_msg()
<< " 2: Input bf16, Weight fp32, Output bf16\n"
<< " 3: Input fp16, Weight fp16, Output fp16, Gemm bf8@fp8\n"
<< " 4: Input int8, Weight int8, Output int8\n"
<< " 5: Input bf16, Weight bf16, Output bf16)\n"
<< " 5: Input bf16, Weight bf16, Output bf16\n"
<< " 6: Input fp32, Weight fp32, Output fp32, Compute tf32)\n"
<< "arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, "
"N, K, Ho, Wo]\n"
<< " 1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, "
@@ -97,6 +99,9 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
using BF16 = ck::bhalf_t;
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
#if defined(__gfx942__)
using TF32 = ck::tf32_t;
#endif
using namespace ck::tensor_layout::convolution;
@@ -155,6 +160,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
@@ -171,6 +182,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
@@ -191,6 +208,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NGCHW_GKYXC_NGKHW)
{
@@ -218,6 +241,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
{
return profile(I2, NGCHW{}, GKCYX{}, NGKHW{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NGCHW{}, GKCYX{}, NGKHW{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
@@ -239,6 +268,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
return profile(
I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, int8_t{}, int8_t{}, int8_t{}, int8_t{}, int8_t{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
@@ -269,6 +304,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
return profile(
I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, int8_t{}, int8_t{}, int8_t{}, int8_t{}, int8_t{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NGCHW_GKYXC_NGKHW)
{
@@ -297,6 +338,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
{
return profile(I3, NGCDHW{}, GKCZYX{}, NGKDHW{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, NGCDHW{}, GKCZYX{}, NGKDHW{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
std::cout << "this data_type & layout is not implemented" << std::endl;

36
profiler/src/profile_grouped_conv_fwd.cpp
View File

@@ -226,6 +226,12 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
{
return profile(I1, GNWC{}, GKXC{}, GNWK{}, INT8{}, INT8{}, INT8{}, INT8{}, INT8{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
@@ -245,6 +251,12 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
{
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, INT8{}, INT8{}, INT8{}, INT8{}, INT8{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
@@ -292,6 +304,12 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
{
return profile(I1, NWGC{}, GKXC{}, NWGK{}, INT8{}, INT8{}, INT8{}, INT8{}, INT8{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I1, NWGC{}, GKXC{}, NWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
@@ -311,6 +329,12 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, INT8{}, INT8{}, INT8{}, INT8{}, INT8{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NGCHW_GKYXC_NGKHW)
{
@@ -326,6 +350,12 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
{
return profile(I2, NGCHW{}, GKYXC{}, NGKHW{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NGCHW{}, GKYXC{}, NGKHW{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NGCHW_GKCYX_NGKHW)
{
@@ -341,6 +371,12 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
{
return profile(I2, NGCHW{}, GKCYX{}, NGKHW{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NGCHW{}, GKCYX{}, NGKHW{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{

35
profiler/src/profile_grouped_conv_fwd_bias_clamp.cpp
View File

@@ -20,14 +20,15 @@ enum struct ConvLayout
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F8_F8, // 4
BF8_BF8_F8, // 5
F8_BF8_F8, // 6
BF8_F8_F8, // 7
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F8_F8, // 4
BF8_BF8_F8, // 5
F8_BF8_F8, // 6
BF8_F8_F8, // 7
F32_F32_F32_TF32, // 8
};
enum struct IndexType
@@ -51,7 +52,8 @@ static void print_helper_msg()
<< " 4: Input fp8, Weight fp8, Output fp8\n"
<< " 5: Input bf8, Weight bf8, Output fp8\n"
<< " 6: Input fp8, Weight bf8, Output fp8\n"
<< " 7: Input bf8, Weight fp8, Output fp8)\n"
<< " 7: Input bf8, Weight fp8, Output fp8\n"
<< " 8: Input fp32, Weight fp32, Output fp32, Compute tf32)\n"
<< "arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
<< " 1: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K]\n"
<< " 2: Input[N, G, C, Hi, Wi], Weight[G, K, Y, X, C], Output[N, "
@@ -103,6 +105,9 @@ int grouped_conv_fwd_bias_clamp(int argc, char* argv[])
using F32 = float;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
#if defined(__gfx942__)
using TF32 = ck::tf32_t;
#endif
using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
@@ -165,6 +170,12 @@ int grouped_conv_fwd_bias_clamp(int argc, char* argv[])
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
@@ -181,6 +192,12 @@ int grouped_conv_fwd_bias_clamp(int argc, char* argv[])
return profile(
I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
std::cout << "this data_type & layout is not implemented" << std::endl;

35
profiler/src/profile_grouped_conv_fwd_clamp.cpp
View File

@@ -20,14 +20,15 @@ enum struct ConvLayout
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F8_F8, // 4
BF8_BF8_F8, // 5
F8_BF8_F8, // 6
BF8_F8_F8, // 7
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F8_F8, // 4
BF8_BF8_F8, // 5
F8_BF8_F8, // 6
BF8_F8_F8, // 7
F32_F32_F32_TF32, // 8
};
enum struct IndexType
@@ -51,7 +52,8 @@ static void print_helper_msg()
<< " 4: Input fp8, Weight fp8, Output fp8\n"
<< " 5: Input bf8, Weight bf8, Output fp8\n"
<< " 6: Input fp8, Weight bf8, Output fp8\n"
<< " 7: Input bf8, Weight fp8, Output fp8)\n"
<< " 7: Input bf8, Weight fp8, Output fp8\n"
<< " 8: Input fp32, Weight fp32, Output fp32, Compute tf32)\n"
<< "arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
<< " 1: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K]\n"
<< " 2: Input[N, G, C, Hi, Wi], Weight[G, K, Y, X, C], Output[N, "
@@ -103,6 +105,9 @@ int grouped_conv_fwd_clamp(int argc, char* argv[])
using F32 = float;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
#if defined(__gfx942__)
using TF32 = ck::tf32_t;
#endif
using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
@@ -168,6 +173,12 @@ int grouped_conv_fwd_clamp(int argc, char* argv[])
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
@@ -184,6 +195,12 @@ int grouped_conv_fwd_clamp(int argc, char* argv[])
return profile(
I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F32_F32_F32_TF32)
{
#if defined(__gfx942__)
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{}, TF32{}, TF32{});
#endif
}
}
std::cout << "this data_type & layout is not implemented" << std::endl;