Skip to content

GitLab

Commit 2f88070a authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into e2e_kernellib
parents 0bb08f4b b076a02a
Hide whitespace changes
Inline Side-by-side
Showing with 299 additions and 180 deletions
Dockerfile
View file @ 2f88070a
FROM ubuntu:20.04 FROM ubuntu:20.04
ARG ROCMVERSION=5.4.3 ARG ROCMVERSION=5.6
ARG compiler_version="release" ARG compiler_version=""
ARG compiler_commit="" ARG compiler_commit=""
RUN set -xe RUN set -xe
...@@ -11,14 +11,14 @@ RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins ...@@ -11,14 +11,14 @@ RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
# Add rocm repository # Add rocm repository
RUN apt-get update RUN apt-get update
RUN apt-get install -y wget gnupg curl RUN apt-get install -y wget gnupg curl
RUN --mount=type=ssh if [ "$ROCMVERSION" != "5.5"]; then \ RUN --mount=type=ssh if [ "$ROCMVERSION" != "5.6"]; then \
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - ; \ wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"; \
else sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \ else sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amd-nonfree-radeon_20.04-1_all.deb" && \
apt update && apt-get install -y ./amd-nonfree-radeon_20.04-1_all.deb && \ apt update && apt-get install -y ./amd-nonfree-radeon_20.04-1_all.deb && \
sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.5 rel-50 > /etc/apt/sources.list.d/rocm-build.list' && \ amdgpu-repo --amdgpu-build=1567752 --rocm-build=compute-rocm-dkms-no-npi-hipclang/11914 && \
amdgpu-repo --amdgpu-build=1558725 && DEBIAN_FRONTEND=noninteractive amdgpu-install -y --usecase=rocm ; \ DEBIAN_FRONTEND=noninteractive amdgpu-install -y --usecase=rocm ; \
fi fi
RUN sh -c "echo deb [arch=amd64] $DEB_ROCM_REPO ubuntu main > /etc/apt/sources.list.d/rocm.list"
RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add - RUN wget --no-check-certificate -qO - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | apt-key add -
RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
...@@ -103,7 +103,7 @@ ENV compiler_commit=$compiler_commit ...@@ -103,7 +103,7 @@ ENV compiler_commit=$compiler_commit
RUN sh -c "echo compiler version = '$compiler_version'" RUN sh -c "echo compiler version = '$compiler_version'"
RUN sh -c "echo compiler commit = '$compiler_commit'" RUN sh -c "echo compiler commit = '$compiler_commit'"
RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_version" !=~ ^"rc" ] && [ "$compiler_commit" = "" ]; then \ RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \ git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && mkdir build && cd build && \ cd llvm-project && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \ cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
...@@ -111,7 +111,7 @@ RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_ver ...@@ -111,7 +111,7 @@ RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_ver
else echo "using the release compiler"; \ else echo "using the release compiler"; \
fi fi
RUN --mount=type=ssh if [ "$compiler_version" != "release" ] && [ "$compiler_version" !=~ ^"rc" ] && [ "$compiler_commit" != "" ]; then \ RUN --mount=type=ssh if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \
git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \ git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \ cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \
cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \ cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
......
Jenkinsfile
View file @ 2f88070a
...@@ -19,22 +19,32 @@ def runShell(String command){ ...@@ -19,22 +19,32 @@ def runShell(String command){
def getDockerImageName(){ def getDockerImageName(){
def img def img
if (params.ROCMVERSION != "5.5"){ if (params.ROCMVERSION != "5.6"){
if (params.COMPILER_COMMIT == ""){ if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}" img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
} }
else{ else{
def commit = "${params.COMPILER_COMMIT}"[0..6] if (params.COMPILER_COMMIT == ""){
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}" img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
}
else{
def commit = "${params.COMPILER_COMMIT}"[0..6]
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
}
} }
} }
else{ else{
if (params.COMPILER_COMMIT == ""){ if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}" img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}"
} }
else{ else{
def commit = "${params.COMPILER_COMMIT}"[0..6] if (params.COMPILER_COMMIT == ""){
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}" img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
}
else{
def commit = "${params.COMPILER_COMMIT}"[0..6]
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
}
} }
} }
return img return img
...@@ -587,7 +597,7 @@ def process_results(Map conf=[:]){ ...@@ -587,7 +597,7 @@ def process_results(Map conf=[:]){
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true
0 21 * * * % COMPILER_VERSION=release;COMPILER_COMMIT= 0 21 * * * % ROCMVERSION=5.5;COMPILER_VERSION=release;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : "" 0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
pipeline { pipeline {
...@@ -605,16 +615,16 @@ pipeline { ...@@ -605,16 +615,16 @@ pipeline {
description: "Force building docker image (default: false), set to true if docker image needs to be updated.") description: "Force building docker image (default: false), set to true if docker image needs to be updated.")
string( string(
name: 'ROCMVERSION', name: 'ROCMVERSION',
defaultValue: '5.4.3', defaultValue: '5.6',
description: 'Specify which ROCM version to use: 5.4.3 (default).') description: 'Specify which ROCM version to use: 5.6 (default).')
string( string(
name: 'COMPILER_VERSION', name: 'COMPILER_VERSION',
defaultValue: 'amd-stg-open', defaultValue: '',
description: 'Specify which version of compiler to use: ck-9110, release, or amd-stg-open (default).') description: 'Specify which version of compiler to use: release, amd-stg-open, or leave blank (default).')
string( string(
name: 'COMPILER_COMMIT', name: 'COMPILER_COMMIT',
defaultValue: '5541927df00eabd6a110180170eca7785d436ee3', defaultValue: '',
description: 'Specify which commit of compiler branch to use: leave empty to use the latest commit, or use 5541927df00eabd6a110180170eca7785d436ee3 (default) commit of amd-stg-open branch.') description: 'Specify which commit of compiler branch to use: leave blank to use the latest commit, or use 5541927df00eabd6a110180170eca7785d436ee3 (default) commit of amd-stg-open branch.')
string( string(
name: 'BUILD_COMPILER', name: 'BUILD_COMPILER',
defaultValue: 'hipcc', defaultValue: 'hipcc',
......
client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
View file @ 2f88070a
...@@ -17,22 +17,22 @@ using InDataType = ck::half_t; ...@@ -17,22 +17,22 @@ using InDataType = ck::half_t;
using WeiDataType = ck::half_t; using WeiDataType = ck::half_t;
using OutDataType = 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 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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 32; static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256; static constexpr ck::index_t N = 256; // batch size
static constexpr ck::index_t K = 192; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 192; static constexpr ck::index_t C = 32; // input channel (per group)
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 28; static constexpr ck::index_t Hi = 28; // input H
static constexpr ck::index_t Wi = 28; static constexpr ck::index_t Wi = 28; // input W
static constexpr ck::index_t Ho = 28; static constexpr ck::index_t Ho = 28; // output H
static constexpr ck::index_t Wo = 28; static constexpr ck::index_t Wo = 28; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -52,50 +52,24 @@ struct SimpleDeviceMem ...@@ -52,50 +52,24 @@ struct SimpleDeviceMem
int main() int main()
{ {
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C}; // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1}; // 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, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C}; std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1}; 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, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K}; std::array<ck::index_t, 5> wei_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1}; 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::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));
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1}; 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> filter_dilations{1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1}; std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_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 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, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout, InLayout,
...@@ -155,9 +129,9 @@ int main() ...@@ -155,9 +129,9 @@ int main()
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); 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 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(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 tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
......
client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
View file @ 2f88070a
...@@ -17,26 +17,26 @@ using BiasDataType = int32_t; ...@@ -17,26 +17,26 @@ using BiasDataType = int32_t;
using RequantScaleDataType = float; using RequantScaleDataType = float;
using OutDataType = 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 WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_K; using BiasLayout = ck::tensor_layout::convolution::G_K;
using RequantScaleLayout = 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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu; using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>; using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2; 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 N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel static constexpr ck::index_t K = 32; // output channel
static constexpr ck::index_t C = 192; // input 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 Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
SimpleDeviceMem() = delete; SimpleDeviceMem() = delete;
...@@ -55,8 +55,11 @@ struct SimpleDeviceMem ...@@ -55,8 +55,11 @@ struct SimpleDeviceMem
int main(int argc, char* argv[]) 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_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_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> 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_lengths{G, N, K, Ho, Wo};
...@@ -64,17 +67,18 @@ int main(int argc, char* argv[]) ...@@ -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_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> 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_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_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_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_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1}; std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C); SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C); SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
NumDimSpatial, NumDimSpatial,
......
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ 2f88070a
...@@ -16,19 +16,19 @@ using WeiDataType = int8_t; ...@@ -16,19 +16,19 @@ using WeiDataType = int8_t;
using BiasDataType = int32_t; using BiasDataType = int32_t;
using OutDataType = 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 WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_K; 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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu; using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>; using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2; 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 N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel static constexpr ck::index_t K = 32; // output channel
static constexpr ck::index_t C = 192; // input 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 Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H static constexpr ck::index_t Hi = 71; // input H
...@@ -55,23 +55,27 @@ struct SimpleDeviceMem ...@@ -55,23 +55,27 @@ struct SimpleDeviceMem
int main(int argc, char* argv[]) 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_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_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> 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_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> 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_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_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_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_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1}; std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C); SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C); SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
View file @ 2f88070a
...@@ -17,21 +17,21 @@ using BiasDataType = int32_t; ...@@ -17,21 +17,21 @@ using BiasDataType = int32_t;
using RequantScaleDataType = float; using RequantScaleDataType = float;
using OutDataType = 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 WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_K; using BiasLayout = ck::tensor_layout::convolution::G_K;
using RequantScaleLayout = 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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::TanH; using ActivationOp = ck::tensor_operation::element_wise::TanH;
using OutElementOp = using OutElementOp =
ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>; ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2; 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 N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel static constexpr ck::index_t K = 32; // output channel
static constexpr ck::index_t C = 192; // input 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 Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H static constexpr ck::index_t Hi = 71; // input H
...@@ -58,8 +58,11 @@ struct SimpleDeviceMem ...@@ -58,8 +58,11 @@ struct SimpleDeviceMem
int main(int argc, char* argv[]) 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_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_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> 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_lengths{G, N, K, Ho, Wo};
...@@ -67,17 +70,18 @@ int main(int argc, char* argv[]) ...@@ -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_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> 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_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_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_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_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1}; std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C); SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C); SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
NumDimSpatial, NumDimSpatial,
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
View file @ 2f88070a
...@@ -16,19 +16,19 @@ using WeiDataType = int8_t; ...@@ -16,19 +16,19 @@ using WeiDataType = int8_t;
using BiasDataType = int32_t; using BiasDataType = int32_t;
using OutDataType = 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 WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_K; 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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::TanH; using ActivationOp = ck::tensor_operation::element_wise::TanH;
using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>; 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 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 N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel static constexpr ck::index_t K = 32; // output channel
static constexpr ck::index_t C = 192; // input 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 Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H static constexpr ck::index_t Hi = 71; // input H
...@@ -56,23 +56,27 @@ struct SimpleDeviceMem ...@@ -56,23 +56,27 @@ struct SimpleDeviceMem
int main(int argc, char* argv[]) 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_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_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> 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_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> 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_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_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_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_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1}; std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C); SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C); SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
......
client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
View file @ 2f88070a
...@@ -16,25 +16,25 @@ using WeiDataType = int8_t; ...@@ -16,25 +16,25 @@ using WeiDataType = int8_t;
using RequantScaleDataType = float; using RequantScaleDataType = float;
using OutDataType = 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 WeiLayout = ck::tensor_layout::convolution::GKYXC;
using RequantScaleLayout = 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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = PassThrough; using ActivationOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>; using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2; 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 N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel static constexpr ck::index_t K = 32; // output channel
static constexpr ck::index_t C = 192; // input 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 Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -54,23 +54,27 @@ struct SimpleDeviceMem ...@@ -54,23 +54,27 @@ struct SimpleDeviceMem
int main(int argc, char* argv[]) 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_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_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> 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_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> 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_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_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_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_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1}; std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C); SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K); SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
......
client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ 2f88070a
...@@ -15,18 +15,18 @@ using InDataType = int8_t; ...@@ -15,18 +15,18 @@ using InDataType = int8_t;
using WeiDataType = int8_t; using WeiDataType = int8_t;
using OutDataType = 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 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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = PassThrough; using ActivationOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>; using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2; 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 N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel static constexpr ck::index_t K = 32; // output channel
static constexpr ck::index_t C = 192; // input 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 Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H static constexpr ck::index_t Hi = 71; // input H
...@@ -53,20 +53,24 @@ struct SimpleDeviceMem ...@@ -53,20 +53,24 @@ struct SimpleDeviceMem
int main(int argc, char* argv[]) 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_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_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> 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_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_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_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_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1}; std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C); SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C); SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K); SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout, InLayout,
......
cmake/EnableCompilerWarnings.cmake
View file @ 2f88070a
...@@ -92,6 +92,7 @@ else() ...@@ -92,6 +92,7 @@ else()
-Wno-unused-command-line-argument -Wno-unused-command-line-argument
-Wno-weak-vtables -Wno-weak-vtables
-Wno-covered-switch-default -Wno-covered-switch-default
-Wno-unsafe-buffer-usage
) )
else() else()
if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "GNU" AND ${COMPILER} MATCHES "CXX") if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "GNU" AND ${COMPILER} MATCHES "CXX")
......
example/15_grouped_gemm/CMakeLists.txt
View file @ 2f88070a
...@@ -5,6 +5,7 @@ add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp) ...@@ -5,6 +5,7 @@ add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp) add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp) add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp) add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
add_dependencies(example_grouped_gemm_xdl add_dependencies(example_grouped_gemm_xdl
...@@ -12,7 +13,8 @@ add_dependencies(example_grouped_gemm_xdl ...@@ -12,7 +13,8 @@ add_dependencies(example_grouped_gemm_xdl
example_grouped_gemm_xdl_fp16 example_grouped_gemm_xdl_fp16
example_grouped_gemm_xdl_bfp16 example_grouped_gemm_xdl_bfp16
example_grouped_gemm_xdl_int8 example_grouped_gemm_xdl_int8
example_grouped_gemm_multiple_d_dl_fp16) example_grouped_gemm_multiple_d_dl_fp16
example_grouped_gemm_xdl_splitk_fp16)
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp) add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
......
example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp 0 → 100644
View file @ 2f88070a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdlSplitKCShuffle
// clang-format off
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
problem_size.Ms = {
167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ns.push_back(768);
problem_size.Ks.push_back(4608);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
}
if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
return !run_grouped_gemm(problem_size, config);
}
example/15_grouped_gemm/run_grouped_gemm_example.inc
View file @ 2f88070a
...@@ -147,6 +147,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co ...@@ -147,6 +147,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
#else #else
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data()); a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data()); b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
c_tensors_device[i]->SetZero();
#endif #endif
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer()); p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
......
example/31_batched_gemm_gemm/CMakeLists.txt
View file @ 2f88070a
add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp) add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp) add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp) add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp) if(NOT GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
endif()
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp) add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
......
example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
View file @ 2f88070a
...@@ -190,11 +190,11 @@ int run_conv2d_fwd_bias_perchannel_quantization_example(const OutElementOp& out_ ...@@ -190,11 +190,11 @@ int run_conv2d_fwd_bias_perchannel_quantization_example(const OutElementOp& out_
const auto in_element_op = InElementOp{}; const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{}; const auto wei_element_op = WeiElementOp{};
using InLayout = ck::tensor_layout::convolution::GNHWC; using InLayout = ck::tensor_layout::convolution::NHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC; using WeiLayout = ck::tensor_layout::convolution::KYXGC;
using BiasLayout = ck::tensor_layout::convolution::G_K; using BiasLayout = ck::tensor_layout::convolution::G_K;
using RequantScaleLayout = 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;
const auto in_g_n_c_wis_desc = const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param); ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
......
example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
View file @ 2f88070a
...@@ -178,10 +178,10 @@ int run_conv2d_fwd_bias_perlayer_quantization_example(const OutElementOp& out_el ...@@ -178,10 +178,10 @@ int run_conv2d_fwd_bias_perlayer_quantization_example(const OutElementOp& out_el
const auto in_element_op = InElementOp{}; const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{}; const auto wei_element_op = WeiElementOp{};
using InLayout = ck::tensor_layout::convolution::GNHWC; using InLayout = ck::tensor_layout::convolution::NHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC; using WeiLayout = ck::tensor_layout::convolution::KYXGC;
using BiasLayout = ck::tensor_layout::convolution::G_K; using BiasLayout = ck::tensor_layout::convolution::G_K;
using OutLayout = ck::tensor_layout::convolution::GNHWK; using OutLayout = ck::tensor_layout::convolution::NHWGK;
const auto in_g_n_c_wis_desc = const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param); ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
......
example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
View file @ 2f88070a
...@@ -180,10 +180,10 @@ int run_conv2d_fwd_perchannel_quantization_example(const OutElementOp& out_eleme ...@@ -180,10 +180,10 @@ int run_conv2d_fwd_perchannel_quantization_example(const OutElementOp& out_eleme
const auto in_element_op = InElementOp{}; const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{}; const auto wei_element_op = WeiElementOp{};
using InLayout = ck::tensor_layout::convolution::GNHWC; using InLayout = ck::tensor_layout::convolution::NHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC; using WeiLayout = ck::tensor_layout::convolution::KYXGC;
using RequantScaleLayout = 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;
const auto in_g_n_c_wis_desc = const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param); ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
......
example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
View file @ 2f88070a
...@@ -162,9 +162,9 @@ int run_conv2d_fwd_perlayer_quantization_example(const OutElementOp& out_element ...@@ -162,9 +162,9 @@ int run_conv2d_fwd_perlayer_quantization_example(const OutElementOp& out_element
const auto in_element_op = InElementOp{}; const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{}; const auto wei_element_op = WeiElementOp{};
using InLayout = ck::tensor_layout::convolution::GNHWC; using InLayout = ck::tensor_layout::convolution::NHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC; using WeiLayout = ck::tensor_layout::convolution::KYXGC;
using OutLayout = ck::tensor_layout::convolution::GNHWK; using OutLayout = ck::tensor_layout::convolution::NHWGK;
const auto in_g_n_c_wis_desc = const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param); ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
......
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View file @ 2f88070a
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp) if(NOT GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
endif()
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
include/ck/ck.hpp
View file @ 2f88070a
...@@ -31,7 +31,7 @@ ...@@ -31,7 +31,7 @@
#ifndef __HIP_DEVICE_COMPILE__ // for host code #ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1 #define CK_BUFFER_RESOURCE_3RD_DWORD -1
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \ #elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) // for GPU code defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#elif defined(__gfx1030__) // for GPU code #elif defined(__gfx1030__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
...@@ -43,8 +43,8 @@ ...@@ -43,8 +43,8 @@
#ifndef __HIP_DEVICE_COMPILE__ // for host code, define nothing #ifndef __HIP_DEVICE_COMPILE__ // for host code, define nothing
#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code #elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
#define CK_USE_AMD_V_MAC_F32 #define CK_USE_AMD_V_MAC_F32
#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || \ #elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \
defined(__gfx1030__) // for GPU code defined(__gfx940__) // for GPU code
#define CK_USE_AMD_V_FMAC_F32 #define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16 #define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8 #define CK_USE_AMD_V_DOT4_I32_I8
...@@ -53,14 +53,18 @@ ...@@ -53,14 +53,18 @@
// MFMA instruction // MFMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code #ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_MFMA #define CK_USE_AMD_MFMA
#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code #elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#define CK_USE_AMD_MFMA #define CK_USE_AMD_MFMA
#endif #endif
#if defined(__gfx90a__) #if(defined(__gfx90a__) || defined(__gfx940__))
#define CK_USE_AMD_MFMA_BF16_1K_OP #define CK_USE_AMD_MFMA_BF16_1K_OP
#endif #endif
#if defined(__gfx940__)
#define CK_USE_AMD_MFMA_GFX940
#endif
// WMMA instruction // WMMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code #ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_WMMA #define CK_USE_AMD_WMMA
...@@ -81,13 +85,13 @@ ...@@ -81,13 +85,13 @@
// buffer atomic add: floating point // buffer atomic add: floating point
#ifndef __HIP_DEVICE_COMPILE__ // for host code #ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code #elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#else // for GPU code #else // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
#endif #endif
#if defined(__gfx90a__) // for GPU code #if(defined(__gfx90a__) || defined(__gfx940__)) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
#else #else
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
...@@ -172,7 +176,10 @@ ...@@ -172,7 +176,10 @@
// denorm test fix, required to work around dissue // denorm test fix, required to work around dissue
#ifndef CK_WORKAROUND_DENORM_FIX #ifndef CK_WORKAROUND_DENORM_FIX
#define CK_WORKAROUND_DENORM_FIX 0 #define CK_WORKAROUND_DENORM_FIX 0
#endif #elif
// enable only on MI200
#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
#endif // CK_WORKAROUND_DENORM_FIX
namespace ck { namespace ck {
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment