mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-17 11:30:02 +00:00
Revise layout of group convolution (#675)
* [What] Remove pure conv int8 instance
[Why] We will never use pure int8 conv in AI, use int8 quantization instead
* Change layout
* Share the kernel parameter
* Support more type of NHWGC for group conv
* Revise client example of conv 2d, use NHWGC layout
* Add instance to cmake
* Revise layout of group conv quantization instance
* Revise layout of external api of group conv quantization
* Revise layout of group conv quantization client example
* Fix clang format
* Add comment to describe meaning of each parameter
[ROCm/composable_kernel commit: 3eecbfb6ec]
This commit is contained in:
rocking
@@ -17,22 +17,22 @@ using InDataType = ck::half_t;
|
||||
using WeiDataType = ck::half_t;
|
||||
using OutDataType = ck::half_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 32;
|
||||
static constexpr ck::index_t N = 256;
|
||||
static constexpr ck::index_t K = 192;
|
||||
static constexpr ck::index_t C = 192;
|
||||
static constexpr ck::index_t Y = 3;
|
||||
static constexpr ck::index_t X = 3;
|
||||
static constexpr ck::index_t Hi = 28;
|
||||
static constexpr ck::index_t Wi = 28;
|
||||
static constexpr ck::index_t Ho = 28;
|
||||
static constexpr ck::index_t Wo = 28;
|
||||
static constexpr ck::index_t N = 256; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 32; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 28; // input H
|
||||
static constexpr ck::index_t Wi = 28; // input W
|
||||
static constexpr ck::index_t Ho = 28; // output H
|
||||
static constexpr ck::index_t Wo = 28; // output W
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
@@ -52,50 +52,24 @@ struct SimpleDeviceMem
|
||||
|
||||
int main()
|
||||
{
|
||||
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
|
||||
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
|
||||
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
|
||||
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1};
|
||||
|
||||
std::partial_sum(rbegin(in_lengths),
|
||||
std::prev(rend(in_lengths)),
|
||||
std::next(rbegin(in_strides)),
|
||||
std::multiplies<>{});
|
||||
std::partial_sum(rbegin(wei_lengths),
|
||||
std::prev(rend(wei_lengths)),
|
||||
std::next(rbegin(wei_strides)),
|
||||
std::multiplies<>{});
|
||||
std::partial_sum(rbegin(out_lengths),
|
||||
std::prev(rend(out_lengths)),
|
||||
std::next(rbegin(out_strides)),
|
||||
std::multiplies<>{});
|
||||
|
||||
// transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
|
||||
std::rotate(
|
||||
rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
|
||||
std::rotate(
|
||||
rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
|
||||
std::rotate(
|
||||
rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
|
||||
std::rotate(
|
||||
rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
|
||||
std::rotate(
|
||||
rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
|
||||
std::rotate(
|
||||
rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> wei_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> wei_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
|
||||
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
|
||||
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
|
||||
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
InLayout,
|
||||
@@ -155,9 +129,9 @@ int main()
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
|
||||
std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
|
||||
sizeof(WeiDataType) * G * K * Y * X * C +
|
||||
sizeof(OutDataType) * G * N * Ho * Wo * K;
|
||||
sizeof(OutDataType) * N * Ho * Wo * G * K;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
|
||||
@@ -17,26 +17,26 @@ using BiasDataType = int32_t;
|
||||
using RequantScaleDataType = float;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::Relu;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
@@ -55,8 +55,11 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
@@ -64,17 +67,18 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
|
||||
NumDimSpatial,
|
||||
|
||||
@@ -16,19 +16,19 @@ using WeiDataType = int8_t;
|
||||
using BiasDataType = int32_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::Relu;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -55,23 +55,27 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp =
|
||||
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
|
||||
@@ -17,21 +17,21 @@ using BiasDataType = int32_t;
|
||||
using RequantScaleDataType = float;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::TanH;
|
||||
using OutElementOp =
|
||||
ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -58,8 +58,11 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
@@ -67,17 +70,18 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
|
||||
NumDimSpatial,
|
||||
|
||||
@@ -16,19 +16,19 @@ using WeiDataType = int8_t;
|
||||
using BiasDataType = int32_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::TanH;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -56,23 +56,27 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp =
|
||||
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
|
||||
@@ -16,25 +16,25 @@ using WeiDataType = int8_t;
|
||||
using RequantScaleDataType = float;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = PassThrough;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
@@ -54,23 +54,27 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp =
|
||||
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
|
||||
@@ -15,18 +15,18 @@ using InDataType = int8_t;
|
||||
using WeiDataType = int8_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = PassThrough;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -53,20 +53,24 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
InLayout,
|
||||
|
||||
Reference in New Issue
Block a user