Skip to content

GitLab

Commit 4100d1d8 authored by Alan Turner's avatar Alan Turner
Browse files

Merge remote-tracking branch 'origin/develop' into migx-flash-attn

parents 48717006 c8a8385f
Hide whitespace changes
Inline Side-by-side
Showing with 236 additions and 359 deletions
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_f16_f16_rank4_reduce4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 4>>& instances)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 4>{});
add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 4>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 4>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_f32_f32_rank3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances)
{
add_device_softmax_f32_f32_rank3_reduce1_instances(instances);
add_device_softmax_f32_f32_rank3_reduce2_instances(instances);
add_device_softmax_f32_f32_rank3_reduce3_instances(instances);
}
void add_device_softmax_f32_f32_rank4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances)
{
add_device_softmax_f32_f32_rank4_reduce1_instances(instances);
add_device_softmax_f32_f32_rank4_reduce2_instances(instances);
add_device_softmax_f32_f32_rank4_reduce3_instances(instances);
add_device_softmax_f32_f32_rank4_reduce4_instances(instances);
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 3;
void add_device_softmax_f32_f32_rank3_reduce1_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 1>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 1>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 3;
void add_device_softmax_f32_f32_rank3_reduce2_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 2>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 2>{});
add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 2>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 2>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 3;
void add_device_softmax_f32_f32_rank3_reduce3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 3>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 3>{});
add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 3>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 3>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_f32_f32_rank4_reduce1_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 1>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_f32_f32_rank4_reduce2_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 2>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 2>{});
add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 2>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 2>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_f32_f32_rank4_reduce3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 3>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 3>{});
add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 3>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 3>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
View file @ 4100d1d8
......@@ -13,12 +13,11 @@ namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_f32_f32_rank4_reduce4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 4>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 4>{});
add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 4>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 4>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_i8_i8_rank3_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances)
{
add_device_softmax_i8_i8_rank3_reduce1_instances(instances);
add_device_softmax_i8_i8_rank3_reduce2_instances(instances);
add_device_softmax_i8_i8_rank3_reduce3_instances(instances);
}
void add_device_softmax_i8_i8_rank4_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances)
{
add_device_softmax_i8_i8_rank4_reduce1_instances(instances);
add_device_softmax_i8_i8_rank4_reduce2_instances(instances);
add_device_softmax_i8_i8_rank4_reduce3_instances(instances);
add_device_softmax_i8_i8_rank4_reduce4_instances(instances);
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 3;
void add_device_softmax_i8_i8_rank3_reduce1_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
{
add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 3;
void add_device_softmax_i8_i8_rank3_reduce2_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
{
add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 2>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 3;
void add_device_softmax_i8_i8_rank3_reduce3_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
{
add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_i8_i8_rank4_reduce1_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
{
add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_i8_i8_rank4_reduce2_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
{
add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 2>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_i8_i8_rank4_reduce3_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
{
add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.cpp deleted 100644 → 0
View file @ 48717006
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <vector>
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp"
#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr index_t RANK = 4;
void add_device_softmax_i8_i8_rank4_reduce4_instances(
std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
{
add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 4>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/utility/device_memory.cpp
View file @ 4100d1d8
......@@ -10,20 +10,57 @@ DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
}
void DeviceMem::Realloc(std::size_t mem_size)
{
if(mpDeviceBuf)
{
hip_check_error(hipFree(mpDeviceBuf));
}
mMemSize = mem_size;
hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
}
void* DeviceMem::GetDeviceBuffer() const { return mpDeviceBuf; }
std::size_t DeviceMem::GetBufferSize() const { return mMemSize; }
void DeviceMem::ToDevice(const void* p) const
{
hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
if(mpDeviceBuf)
{
hip_check_error(
hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
}
else
{
throw std::runtime_error("ToDevice with an empty pointer");
}
}
void DeviceMem::FromDevice(void* p) const
{
hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
if(mpDeviceBuf)
{
hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
}
else
{
throw std::runtime_error("FromDevice with an empty pointer");
}
}
void DeviceMem::SetZero() const { hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize)); }
void DeviceMem::SetZero() const
{
if(mpDeviceBuf)
{
hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize));
}
}
DeviceMem::~DeviceMem() { hip_check_error(hipFree(mpDeviceBuf)); }
DeviceMem::~DeviceMem()
{
if(mpDeviceBuf)
{
hip_check_error(hipFree(mpDeviceBuf));
}
}
profiler/README.md
View file @ 4100d1d8
......@@ -102,4 +102,85 @@ arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
arg.e_grid_desc_m_n_{ 4096, 4096}
....
Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s
## Profile grouped convolution backward data kernels
```bash
# arg1: tensor operation (grouped_conv_bwd_data: Grouped Convolution Backward Data)
# arg2: data type (0: Output fp32, Weight fp32, Input fp32
# 1: Output fp16, Weight fp16, Input fp16
# 2: Output bf16, Weight bf16, Input bf16
# arg3: tensor layout (0: Output[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Input[G, N, Ho, Wo, K]
# 1: Output[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Input[N, Ho, Wo, G, K])
# arg4: verification (0: no, 1: yes)
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
################ op datatype layout verify init log time Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx
./bin/ckProfiler grouped_conv_bwd_data 1 0 1 1 0 1 2 32 4 192 192 3 3 28 28 1 1 1 1 1 1 1 1
```
Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
```
out: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
wei: dim 5, lengths {32, 192, 192, 3, 3}, strides {331776, 1728, 1, 576, 192}
in: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
....
Best configuration parameters:
name: DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<256, 128, 256, 32, 8, 2, Default, 32, 32, 2, 4, 8, 4, 1, 1>
avg_time: 0.768321
tflops: 86.6679
GB/s: 127.947
```
## Profile grouped convolution backward weight kernels
```bash
# arg1: tensor operation (grouped_conv_bwd_weight: Grouped Convolution Backward Weight)
# arg2: data type (0: Input fp32, Weight fp32, Output fp32
# 1: Input fp16, Weight fp16, Output fp16
# 2: Input bf16, Weight fp32, Output bf16)
# arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, N, K, Ho, Wo]
# 1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]
# 2: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K]
# arg4: verification (0: no, 1: yes)
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
# SplitK
################ op datatype layout verify init log time Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx SplitK
./bin/ckProfiler grouped_conv_bwd_weight 1 0 1 1 0 1 2 32 256 256 512 3 3 28 28 1 1 1 1 1 0 0 0 1
```
Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
```
input: dim 5, lengths {32, 512, 1024, 28, 28}, strides {411041792, 802816, 1, 28672, 1024}
weight: dim 5, lengths {32, 512, 1024, 3, 3}, strides {4718592, 9216, 1, 3072, 1024}
output: dim 5, lengths {32, 512, 512, 26, 26}, strides {177209344, 346112, 1, 13312, 512}
....
Best configuration parameters:
name: DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
avg_time: 68.5216
tflops: 95.337
GB/s: 69.2301
```
Note: This kernel use atomic add, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View file @ 4100d1d8
......@@ -94,7 +94,6 @@ bool profile_gemm_splitk_impl(int do_verification,
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.SetZero();
using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
BLayout,
......@@ -136,77 +135,114 @@ bool profile_gemm_splitk_impl(int do_verification,
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
KBatch);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 36, 40, 60,
64, 72, 80, 88, 96, 128, 144, 160, 176, 192, 256};
if(KBatch > 0)
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
kbatch_list = {KBatch};
}
for(std::size_t i = 0; i < kbatch_list.size(); i++)
{
auto kbatch_curr = kbatch_list[i];
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
kbatch_curr);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string op_name = op_ptr->GetTypeString();
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
std::size_t flop = std::size_t(2) * M * N * K;
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
std::string op_name = op_ptr->GetTypeString();
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::size_t flop = std::size_t(2) * M * N * K;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
float gb_per_sec = num_btype / 1.E6 / ave_time;
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
if(do_log)
// set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
is_same_v<CDataType, f8_t>)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
std::string msg = "Error: Incorrect results!";
double rtol = 1e-1;
double atol = 1e-1;
pass = pass & ck::utils::check_err(
c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
}
else
{
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
}
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
}
......@@ -246,7 +282,7 @@ bool profile_gemm_splitk_impl(int do_verification,
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << KBatch
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
......
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