Skip to content

GitLab

Unverified Commit 760ea189 authored by arai713's avatar arai713 Committed by GitHub
Browse files

Merge branch 'develop' into codegen_hiprtc

parents c87aa6c8 cb8c7f42
Show whitespace changes
Inline Side-by-side
Showing with 197 additions and 125 deletions
Dockerfile
View file @ 760ea189
......@@ -116,7 +116,7 @@ ENV compiler_commit=$compiler_commit
RUN sh -c "echo compiler version = '$compiler_version'" && \
sh -c "echo compiler commit = '$compiler_commit'"
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline-open" ] ) && [ "$compiler_commit" = "" ]; then \
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline" ] ) && [ "$compiler_commit" = "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \
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" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
......@@ -124,7 +124,7 @@ RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd
else echo "using the release compiler"; \
fi
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline-open" ] ) && [ "$compiler_commit" != "" ]; then \
RUN if ( [ "$compiler_version" = "amd-staging" ] || [ "$compiler_version" = "amd-mainline" ] ) && [ "$compiler_commit" != "" ]; then \
git clone -b "$compiler_version" https://github.com/ROCm/llvm-project.git && \
cd llvm-project && git checkout "$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" -DLLVM_ENABLE_RUNTIMES="compiler-rt" ../llvm && \
......
Jenkinsfile
View file @ 760ea189
......@@ -133,7 +133,7 @@ def buildDocker(install_prefix){
def image_name = getDockerImageName()
echo "Building Docker for ${image_name}"
def dockerArgs = "--squash --build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg CK_SCCACHE='${env.CK_SCCACHE}' --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' --build-arg DISABLE_CACHE='git rev-parse ${params.COMPILER_VERSION}' "
if(params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
if(params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline" || params.COMPILER_COMMIT != ""){
dockerArgs = dockerArgs + " --no-cache "
}
echo "Build Args: ${dockerArgs}"
......@@ -358,7 +358,7 @@ def buildHipClangJob(Map conf=[:]){
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg CK_SCCACHE='${env.CK_SCCACHE}' --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
def video_id = sh(returnStdout: true, script: 'getent group video | cut -d: -f3')
......@@ -549,7 +549,7 @@ def Build_CK(Map conf=[:]){
dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
}
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline" || params.COMPILER_COMMIT != ""){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
}
if(params.BUILD_LEGACY_OS){
......@@ -737,7 +737,7 @@ def process_results(Map conf=[:]){
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true;RUN_CODEGEN_TESTS=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
0 13 * * * % BUILD_LEGACY_OS=true''' : ""
......@@ -765,7 +765,7 @@ pipeline {
string(
name: 'COMPILER_VERSION',
defaultValue: '',
description: 'Specify which version of compiler to use: release, amd-staging, amd-mainline-open, or leave blank (default).')
description: 'Specify which version of compiler to use: release, amd-staging, amd-mainline, or leave blank (default).')
string(
name: 'COMPILER_COMMIT',
defaultValue: '',
......
example/ck_tile/12_smoothquant/CMakeLists.txt
View file @ 760ea189
......@@ -18,7 +18,7 @@ function (add_smoothquant_example TARGET_NAME MAIN_SRC)
target_compile_options(${TARGET_NAME} PRIVATE ${COMPILE_OPTIONS})
endfunction(add_smoothquant_example TARGET_NAME MAIN_SRC)
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_smoothquant_example(tile_smoothquant smoothquant.cpp ${INSTANCE_SRCS})
add_smoothquant_example(tile_example_smoothquant example_smoothquant.cpp)
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_smoothquant_example(tile_smoothquant smoothquant.cpp ${INSTANCE_SRCS})
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.hpp
View file @ 760ea189
......@@ -269,9 +269,9 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
......@@ -279,7 +279,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
......@@ -341,6 +340,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
......@@ -350,7 +350,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
......@@ -396,6 +395,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
......@@ -405,7 +405,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
......@@ -447,6 +446,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
......@@ -456,7 +456,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
......@@ -760,16 +759,15 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// NOTE: Synchronize threads in a workgroup at the start of each MAC
// cluster, but except the first, as we can shorten non-MAC cluster a bit
......@@ -866,6 +864,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
......@@ -874,7 +873,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
......@@ -942,6 +940,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
......@@ -950,7 +949,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
......@@ -1018,6 +1016,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k0 * KPerInnerLoop>{}),
......@@ -1026,7 +1025,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp
View file @ 760ea189
......@@ -305,6 +305,7 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(I0));
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
......@@ -314,7 +315,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
b_thread_bufs(I0));
});
});
});
// Global prefetch 3
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
......@@ -356,9 +356,9 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(lds_read_buf),
b_thread_desc_,
......@@ -366,7 +366,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
b_thread_bufs(lds_read_reg_buf));
});
});
});
a_blockwise_copy.RunWrite(
a_block_desc, a_block_buf.At(lds_write_buf), vmem_buf);
......@@ -437,6 +436,7 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
......@@ -446,7 +446,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
b_thread_bufs(lds_read_reg_buf));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf), vmem_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf), vmem_buf);
......@@ -496,6 +495,7 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
......@@ -505,7 +505,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
b_thread_bufs(lds_read_reg_buf));
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
......
include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_one_pass.hpp
View file @ 760ea189
......@@ -30,6 +30,7 @@ struct AddRmsnorm2dRdquantFwdPipelineOnePass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockAddRmsnorm2dRdquantFwdProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
......@@ -69,6 +70,13 @@ struct AddRmsnorm2dRdquantFwdPipelineOnePass
auto reduce_square_sum_func = ReduceOp::SquareAdd{};
auto reduce_sum_func = ReduceOp::Add{};
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
......@@ -116,8 +124,23 @@ struct AddRmsnorm2dRdquantFwdPipelineOnePass
});
// compute absmax, each-thread->cross-lane->cross-warp
auto absmax = block_reduce2d(
auto absmax = [&]() {
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
{
return block_reduce2d(y,
reduce_absmax_func.GetIdentityValue<ComputeDataType>(),
reduce_absmax3_func,
sequence<1, 2>{});
}
else
{
return block_reduce2d(
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
}
}();
block_reduce2d_sync(absmax, reduce_max_func);
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
......
include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_pass.hpp
View file @ 760ea189
......@@ -30,6 +30,7 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockAddRmsnorm2dRdquantFwdProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
......@@ -76,6 +77,13 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
auto reduce_square_sum_func = ReduceOp::SquareAdd{};
auto reduce_sum_func = ReduceOp::Add{};
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
......@@ -177,6 +185,12 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
y(idx) = type_convert<ComputeDataType>(y_);
});
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
block_reduce2d(y, absmax, reduce_absmax3_func, sequence<1, 2>{});
else
block_reduce2d(y, absmax, reduce_absmax_func);
if constexpr(kSaveX)
......
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp
View file @ 760ea189
......@@ -25,6 +25,7 @@ struct SmoothquantPipelineOnePass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
......@@ -52,7 +53,15 @@ struct SmoothquantPipelineOnePass
xscale_window_, Policy::template MakeXScaleBlockTileDistribution<Problem>());
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
auto block_reduce2d_cross_warp_sync =
......@@ -68,8 +77,23 @@ struct SmoothquantPipelineOnePass
xscale);
// compute absmax, cross-lane->cross-warp
auto absmax = block_reduce2d(
auto absmax = [&]() {
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
{
return block_reduce2d(y,
reduce_absmax_func.GetIdentityValue<ComputeDataType>(),
reduce_absmax3_func,
sequence<1, 2>{});
}
else
{
return block_reduce2d(
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
}
}();
block_reduce2d_sync(absmax, reduce_max_func);
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
......
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp
View file @ 760ea189
......@@ -25,6 +25,7 @@ struct SmoothquantPipelineTwoPass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
......@@ -56,6 +57,13 @@ struct SmoothquantPipelineTwoPass
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
......@@ -77,6 +85,12 @@ struct SmoothquantPipelineTwoPass
x,
xscale);
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
block_reduce2d(y, absmax, reduce_absmax3_func, sequence<1, 2>{});
else
block_reduce2d(y, absmax, reduce_absmax_func);
move_tile_window(x_window, {0, Block_N});
......
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