Skip to content

GitLab

Commit 4947639c authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop' into amd-master

parents 17cf8179 d39c3f5d
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 6040 additions and 103 deletions
example/ck_tile/01_fmha/generate.py
View file @ 4947639c
This diff is collapsed.
example/ck_tile/01_fmha/script/benchmark_bwd.sh 0 → 100644
View file @ 4947639c
#!/bin/sh
# TODO: run this script from CK root
BUILD=build
EXE=$BUILD/bin/tile_example_fmha_bwd
VALID=0
for prec in "fp16" "bf16" ; do
for perm in 0 1 ; do
for hdim in 32 64 128 ; do
nhead=$((2048 / $hdim)) # follow fav2 setup
$EXE -prec=$prec -b=32 -h=$nhead -d=$hdim -s=512 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=16 -h=$nhead -d=$hdim -s=1024 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=8 -h=$nhead -d=$hdim -s=2048 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=4 -h=$nhead -d=$hdim -s=4096 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=2 -h=$nhead -d=$hdim -s=8192 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=1 -h=$nhead -d=$hdim -s=16384 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
done
done
done
example/ck_tile/01_fmha/script/smoke_test_bwd.sh 0 → 100644
View file @ 4947639c
#!/bin/sh
# TODO: run this script from CK root
BUILD=build
EXE=$BUILD/bin/tile_example_fmha_bwd
KNAME=1
export CK_WARMUP=0
export CK_REPEAT=1
COMMON_ARGS='-v=1'
for prec in "fp16" "bf16" ; do
for perm in 0 1 ; do
for hdim in 32 64 128 ; do
for mode in 0 1 ; do
for bias in "n" "e" "a"; do
for dbias in 0 1 ; do
for p_drop in 0.0 0.2; do
$EXE -prec=$prec -b=1 -h=4 -h_k=2 -d=$hdim -s=259 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=2 -d=$hdim -s=516 -s_k=253 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=4 -h_k=1 -d=$hdim -s=500 -s_k=251 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=1 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=2 -d=$hdim -s=900 -s_k=258 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=2 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=1 -d=$hdim -s=987 -s_k=219 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=t:128,30 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=3 -h_k=1 -d=$hdim -s=244 -s_k=499 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=b:4,35 -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
done
done
done
done
done
done
done
example/ck_tile/01_fmha/script/smoke_test.sh → example/ck_tile/01_fmha/script/smoke_test_fwd.sh
View file @ 4947639c
...@@ -17,17 +17,19 @@ for perm in 0 1 ; do ...@@ -17,17 +17,19 @@ for perm in 0 1 ; do
for vlayout in "r" "c" ; do for vlayout in "r" "c" ; do
for hdim in 32 64 128 256 ; do for hdim in 32 64 128 256 ; do
for lse in 0 1 ; do for lse in 0 1 ; do
for bias in "n" "e" "a"; do for bias in "n" "e" "a" ; do
for p_drop in 0.0 0.2; do
# $EXE -prec=$prec -mode=$mode -b=1 -h=1 -d=$hdim -s=1024 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS # $EXE -prec=$prec -mode=$mode -b=1 -h=1 -d=$hdim -s=1024 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=2 -h_k=1 -d=16, -d_v=$hdim -s=55 -s_k=256 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=2 -h=2 -h_k=1 -d=16, -d_v=$hdim -s=55 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=3 -d=$hdim -s=100 -s_k=51 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=1 -h=3 -d=$hdim -s=100 -s_k=51 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=16 -d_v=$hdim -s=99 -s_k=256 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=1 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=16 -d_v=$hdim -s=99 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=1 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1024 -s_k=256 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1024 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -d_v=24 -s=3 -s_k=99 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -d_v=24 -s=3 -s_k=99 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=3 -h=2 -h_k=1 -d=$hdim -s=200 -s_k=520 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=t:128,30 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=3 -h=2 -h_k=1 -d=$hdim -s=200 -s_k=520 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=t:128,30 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -s=99 -s_k=32 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=b:4,35 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -s=99 -s_k=32 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=b:4,35 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=33 -s_k=0 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS $EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=33 -s_k=0 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1 -s_k=10 -s_kpad=32 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
done done
done done
...@@ -36,6 +38,7 @@ done ...@@ -36,6 +38,7 @@ done
done done
done done
done done
done
for perm in 0 1 ; do for perm in 0 1 ; do
for bias in "n" "e" "a" ; do for bias in "n" "e" "a" ; do
......
example/ck_tile/01_fmha/utils.hpp
View file @ 4947639c
...@@ -4,12 +4,14 @@ ...@@ -4,12 +4,14 @@
#pragma once #pragma once
#include <cstdint> #include <cstdint>
#include <cstdlib>
#include <optional> #include <optional>
#include <ostream> #include <ostream>
#include <tuple> #include <tuple>
#include <utility> #include <utility>
#include <vector> #include <vector>
#include <functional> #include <functional>
#include <string>
#include "ck_tile/core/container/span.hpp" #include "ck_tile/core/container/span.hpp"
...@@ -37,12 +39,14 @@ std::vector<int32_t> to_seqstarts(ck_tile::span<const int32_t> seqlens) ...@@ -37,12 +39,14 @@ std::vector<int32_t> to_seqstarts(ck_tile::span<const int32_t> seqlens)
std::vector<int32_t> generate_seqlens(mode_enum mode, std::vector<int32_t> generate_seqlens(mode_enum mode,
unsigned count, unsigned count,
int32_t seqlens_sum, int32_t seqlen_avg,
int32_t seqlen_max = -1, // if not negative, clamp max
std::optional<unsigned> seed = std::nullopt) std::optional<unsigned> seed = std::nullopt)
{ {
assert(0 < count); assert(0 < count);
std::vector<int32_t> seqlens(count, seqlens_sum); std::vector<int32_t> seqlens(
count, seqlen_max > 0 ? (seqlen_avg < seqlen_max ? seqlen_avg : seqlen_max) : seqlen_avg);
if(mode == mode_enum::group && 1 < count) if(mode == mode_enum::group && 1 < count)
{ {
...@@ -55,7 +59,7 @@ std::vector<int32_t> generate_seqlens(mode_enum mode, ...@@ -55,7 +59,7 @@ std::vector<int32_t> generate_seqlens(mode_enum mode,
std::uniform_int_distribution<size_type> step_dist(1, count - 1); std::uniform_int_distribution<size_type> step_dist(1, count - 1);
auto next_step = std::bind(step_dist, std::ref(random_engine)); auto next_step = std::bind(step_dist, std::ref(random_engine));
for(unsigned repeat = seqlens_sum * (count / 2); 0 < repeat; --repeat) for(unsigned repeat = seqlen_avg * (count / 2); 0 < repeat; --repeat)
{ {
const size_type to_decrease = next_idx(); const size_type to_decrease = next_idx();
// make sure each elements of seqlens is always greater than 0 // make sure each elements of seqlens is always greater than 0
...@@ -66,6 +70,11 @@ std::vector<int32_t> generate_seqlens(mode_enum mode, ...@@ -66,6 +70,11 @@ std::vector<int32_t> generate_seqlens(mode_enum mode,
const size_type to_increase = (to_decrease + next_step()) % count; const size_type to_increase = (to_decrease + next_step()) % count;
if(seqlen_max > 0 && seqlens[to_increase] >= seqlen_max)
{
continue;
}
--seqlens[to_decrease]; --seqlens[to_decrease];
++seqlens[to_increase]; ++seqlens[to_increase];
} }
...@@ -76,10 +85,91 @@ std::vector<int32_t> generate_seqlens(mode_enum mode, ...@@ -76,10 +85,91 @@ std::vector<int32_t> generate_seqlens(mode_enum mode,
std::vector<int32_t> generate_seqstarts(mode_enum mode, std::vector<int32_t> generate_seqstarts(mode_enum mode,
unsigned count, unsigned count,
int32_t seqlens_sum, int32_t seqlen_avg,
int32_t seqlen_max = -1,
std::optional<unsigned> seed = std::nullopt) std::optional<unsigned> seed = std::nullopt)
{ {
return to_seqstarts(generate_seqlens(mode, count, seqlens_sum, seed)); return to_seqstarts(generate_seqlens(mode, count, seqlen_avg, seqlen_max, seed));
}
/*
* decode the seqlen string from cmdline
* example (assume batch=3)
* q_val=1,2,3 k_val=4,5,6 -> OK
* q_val=1,2,3 -> OK, k same as q
* q_val=1,2 -> OK, q will rand remaining 1 element, k same as q
* q_val=1,2 k_val=4,5 -> OK, q/k will rand remaining 1 element
* q_val=1,2,3,4 -> OK, but ignore exceed one
*
* q_val=1,2 k_val=4,5,6 -> not OK, k must have same splits with q
* q_val=1,2 k_val=4 -> not OK, k must have same splits with q
*/
std::tuple<std::vector<ck_tile::index_t>,
std::vector<ck_tile::index_t>,
std::vector<ck_tile::index_t>>
decode_seqlen(mode_enum mode,
ck_tile::index_t batch,
std::string q_val,
std::string k_val,
std::string k_pad_val,
std::optional<unsigned> seed = std::nullopt)
{
#define _S2I_(str_) static_cast<ck_tile::index_t>(std::atoi((str_).c_str()))
if(mode == mode_enum::batch)
{
ck_tile::index_t q = _S2I_(q_val);
ck_tile::index_t k = _S2I_(k_val);
auto s_q = std::vector<ck_tile::index_t>(batch, q);
auto s_k = std::vector<ck_tile::index_t>(batch, k < 0 ? q : k);
auto s_kpad = std::vector<ck_tile::index_t>(batch, -1); // TODO: batch not support k_padding
return std::make_tuple(s_q, s_k, s_kpad);
}
else
{
ck_tile::index_t idx = 0;
std::string::size_type pos_q = 0;
std::string::size_type pos_k = 0;
std::string::size_type pos_kp = 0;
std::vector<ck_tile::index_t> s_q;
std::vector<ck_tile::index_t> s_k;
std::vector<ck_tile::index_t> s_kpad;
while(true)
{
auto found_q = q_val.find(',', pos_q);
auto found_k = k_val.find(',', pos_k);
auto found_kp = k_pad_val.find(',', pos_kp);
ck_tile::index_t q = _S2I_(
q_val.substr(pos_q, found_q == std::string::npos ? found_q : found_q - pos_q));
ck_tile::index_t k = _S2I_(
k_val.substr(pos_k, found_k == std::string::npos ? found_k : found_k - pos_k));
ck_tile::index_t kp = _S2I_(k_pad_val.substr(
pos_kp, found_kp == std::string::npos ? found_kp : found_kp - pos_kp));
s_q.push_back(q);
s_k.push_back(k < 0 ? q : k);
s_kpad.push_back(kp);
idx++;
if(found_q == std::string::npos || idx >= batch)
{
break;
}
pos_q = found_q + 1;
pos_k = found_k == std::string::npos ? pos_k : found_k + 1;
pos_kp = found_kp == std::string::npos ? pos_kp : found_kp + 1;
}
if(idx < batch)
{
auto rem_q = generate_seqlens(mode, batch - idx, s_q.back(), s_kpad.back(), seed);
auto rem_k = generate_seqlens(mode, batch - idx, s_k.back(), s_kpad.back(), seed);
s_q.insert(s_q.end(), rem_q.begin(), rem_q.end());
s_k.insert(s_k.end(), rem_k.begin(), rem_k.end());
s_kpad.insert(s_kpad.end(), batch - idx, s_kpad.back());
}
return std::make_tuple(s_q, s_k, s_kpad);
}
#undef _S2I_
} }
int env_get_int(const char* var_name, int default_int) int env_get_int(const char* var_name, int default_int)
...@@ -87,6 +177,6 @@ int env_get_int(const char* var_name, int default_int) ...@@ -87,6 +177,6 @@ int env_get_int(const char* var_name, int default_int)
char* v = getenv(var_name); char* v = getenv(var_name);
int r = default_int; int r = default_int;
if(v) if(v)
r = atoi(v); r = std::atoi(v);
return r; return r;
} }
include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp 0 → 100644
View file @ 4947639c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3.hpp"
#include "ck/utility/is_detected.hpp"
namespace ck {
// Thread-group level multi-source, multi-destination tensor slice data movement
// Assume:
// 1. All sources and destinations are DynamicBuffer
// 2. Same VectorDim and ScalerPerVector for all sources and destinations
// 3. DstInMemOps are per destination tensor
// 4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
// 5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
//
// Does following things to avoid scratch memory issue
// 1. Pass tensor descritpors by reference (or tuple of references)
// 2. Does not keep reference to tensor descriptor
// 3. Does not construct new tensor coordinate when call Run()
template <typename ThreadGroup,
typename SrcDatas,
typename DstDatas,
typename SrcDescs,
typename DstDescs,
typename ElementwiseOperation,
typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
typename SliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,
typename SrcDimAccessOrder,
typename DstDimAccessOrder,
index_t SrcVectorDim,
index_t DstVectorDim,
typename SrcScalarPerVectors,
index_t DstScalarPerVector,
typename ThreadTransferSrcResetCoordinateAfterRunFlags,
typename ThreadTransferDstResetCoordinateAfterRunFlags,
index_t NumThreadScratch = 1>
struct ThreadGroupTensorSliceTransfer_v7r3
{
static constexpr index_t nDim =
remove_cvref_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
static constexpr index_t nSrc = remove_cvref_t<SrcDescs>::Size();
static constexpr index_t nDst = remove_cvref_t<DstDescs>::Size();
using Index = MultiIndex<nDim>;
static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
__device__ constexpr ThreadGroupTensorSliceTransfer_v7r3(
const SrcDescs& src_descs,
const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
const DstDescs& dst_descs,
const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
const ElementwiseOperation& element_op)
: threadwise_transfer_(src_descs,
StaticallyIndexedArray<Index, nSrc>{},
dst_descs,
StaticallyIndexedArray<Index, nDst>{},
element_op)
{
static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() &&
nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() &&
nDst == DstDatas::Size() && nDst == DstDescs::Size() &&
nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(),
"wrong!");
static_for<0, nSrc, 1>{}([&](auto i) {
static_assert(
nDim == remove_cvref_t<tuple_element_t<i.value, SrcDescs>>::GetNumOfDimension(),
"wrong!");
});
static_for<0, nDst, 1>{}([&](auto i) {
static_assert(
nDim == remove_cvref_t<tuple_element_t<i.value, DstDescs>>::GetNumOfDimension(),
"wrong!");
});
static_assert(nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
"wrong! nDim not consistent");
static_assert(
is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"wrong! ThreadGroup::GetNumOfThread() too small");
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(ThreadGroup::GetThreadId()));
const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
const auto src_thread_slice_origins = generate_tuple(
[&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
Number<nSrc>{});
const auto dst_thread_slice_origins = generate_tuple(
[&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
Number<nDst>{});
threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
}
}
template <typename SrcBuffers, index_t ThreadScratchId = 0>
__device__ void RunRead(const SrcDescs& src_descs,
const SrcBuffers& src_bufs,
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_descs, src_bufs, thread_scratch_id);
}
}
template <typename T>
using is_tuple = decltype(std::declval<T&>().IsTuple());
template <typename DstBuffers, index_t ThreadScratchId = 0>
__device__ void RunWrite(const DstDescs& dst_descs,
DstBuffers dst_bufs,
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
if constexpr(is_detected<is_tuple, decltype(dst_bufs)>::value)
threadwise_transfer_.RunWrite(dst_descs, dst_bufs, thread_scratch_id);
else
threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs), thread_scratch_id);
}
}
template <typename SrcBuffers, typename DstBuffers>
__device__ void Run(const SrcDescs& src_descs,
const SrcBuffers& src_bufs,
const DstDescs& dst_descs,
DstBuffers dst_bufs)
{
RunRead(src_descs, src_bufs);
RunWrite(dst_descs, dst_bufs);
}
template <index_t ISrc>
__device__ void
MoveSrcSliceWindow(const SrcDescs& src_descs, Number<ISrc> iSrc, const Index& step)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step);
}
}
__device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, const Index& step)
{
static_for<0, SrcDescs::Size(), 1>{}(
[&](auto i) { MoveSrcSliceWindow(src_descs, i, step); });
}
template <index_t IDst>
__device__ void
MoveDstSliceWindow(const DstDescs& dst_descs, Number<IDst> iDst, const Index& step)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step);
}
}
__device__ void MoveDstSliceWindow(const DstDescs& dst_descs, const Index& step)
{
static_for<0, DstDescs::Size(), 1>{}(
[&](auto i) { MoveDstSliceWindow(dst_descs, i, step); });
}
private:
static constexpr auto thread_cluster_desc_ =
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v7r3<SrcDatas,
DstDatas,
SrcDescs,
DstDescs,
ElementwiseOperation,
DstInMemOps,
decltype(thread_slice_lengths),
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorDim,
DstVectorDim,
SrcScalarPerVectors,
DstScalarPerVector,
ThreadTransferSrcResetCoordinateAfterRunFlags,
ThreadTransferDstResetCoordinateAfterRunFlags,
NumThreadScratch>;
ThreadwiseTransfer threadwise_transfer_;
};
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp
View file @ 4947639c
...@@ -674,7 +674,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -674,7 +674,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
clear_workspace(); clear_workspace();
}; };
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>( ave_time += ck::utility::launch_and_time_kernel_with_preprocess<false>(
stream_config, stream_config,
run_flush_cache, run_flush_cache,
kernel, kernel,
...@@ -690,7 +690,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle ...@@ -690,7 +690,7 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
} }
else else
{ {
ave_time = launch_and_time_kernel_with_preprocess( ave_time += launch_and_time_kernel_with_preprocess(
stream_config, stream_config,
clear_workspace, clear_workspace,
kernel, kernel,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
View file @ 4947639c
...@@ -820,15 +820,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -820,15 +820,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
return false; return false;
} }
} }
else if(ck::is_lds_direct_load_supported()) if(!ck::is_xdl_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
{
return false;
}
}
else
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 4947639c
...@@ -961,6 +961,29 @@ struct Elu ...@@ -961,6 +961,29 @@ struct Elu
const float alpha_; const float alpha_;
}; };
struct ConvScale
{
__host__ __device__ ConvScale(float scale_in = 1.f,
float scale_wei = 1.f,
float scale_out = 1.f)
: scale_in_(scale_in), scale_wei_(scale_wei), scale_out_(scale_out)
{
}
template <typename E, typename C>
__host__ __device__ void operator()(E& e, const C& c) const;
template <>
__host__ __device__ void operator()<f8_t, float>(f8_t& e, const float& c) const
{
e = type_convert<f8_t>(c * scale_in_ * scale_wei_ * scale_out_);
};
float scale_in_;
float scale_wei_;
float scale_out_;
};
// support fastconvert of int8 to fp16 // support fastconvert of int8 to fp16
template <typename InputDataType, typename OutputDataType, index_t RegPackNumber> template <typename InputDataType, typename OutputDataType, index_t RegPackNumber>
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp
View file @ 4947639c
...@@ -1123,7 +1123,7 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -1123,7 +1123,7 @@ struct GridwiseGemm_xdl_cshuffle_v3
} }
template <typename CGridDesc> template <typename CGridDesc>
__device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( __host__ __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock) const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
{ {
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor( const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
...@@ -1141,26 +1141,22 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -1141,26 +1141,22 @@ struct GridwiseGemm_xdl_cshuffle_v3
using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>; using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
// using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>; // using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>;
template <bool HasMainKBlockLoop, template <typename AGridDesc_AK0_M_K1,
typename BGridDesc_BK0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd> TailNumber TailNum = TailNumber::Odd>
__device__ static void Run(const ADataType* p_a_grid, __device__ static void Run(const ADataType* p_a_grid,
const BDataType* p_b_grid, const BDataType* p_b_grid,
CDataType* p_c_grid, CDataType* p_c_grid,
void* p_shared, void* p_shared,
const Problem& problem) const Problem& problem,
const AGridDesc_AK0_M_K1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1& b_grid_desc_bk0_n_bk1,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock)
{ {
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize()); p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
...@@ -1508,12 +1504,11 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -1508,12 +1504,11 @@ struct GridwiseGemm_xdl_cshuffle_v3
template <bool HasMainKBlockLoop, template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd> TailNumber TailNum = TailNumber::Odd>
__device__ static void Run_2Lds(const ADataType* p_a_grid, __device__ static void Run(const ADataType* p_a_grid,
const BDataType* p_b_grid, const BDataType* p_b_grid,
CDataType* p_c_grid, CDataType* p_c_grid,
void* p_shared_0, void* p_shared,
void* p_shared_1, const Problem& problem)
const Problem& problem)
{ {
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
...@@ -1521,11 +1516,42 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -1521,11 +1516,42 @@ struct GridwiseGemm_xdl_cshuffle_v3
problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0); problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N( const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC); problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock); c_grid_desc_m_n, problem.MBlock, problem.NBlock);
Run<decltype(a_grid_desc_ak0_m_ak1),
decltype(b_grid_desc_bk0_n_bk1),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
HasMainKBlockLoop,
CGlobalMemoryDataOperation,
TailNum>(p_a_grid,
p_b_grid,
p_c_grid,
p_shared,
problem,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock);
}
template <typename AGridDesc_AK0_M_K1,
typename BGridDesc_BK0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run_2Lds(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
void* p_shared_0,
void* p_shared_1,
const Problem& problem,
const AGridDesc_AK0_M_K1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1& b_grid_desc_bk0_n_bk1,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize()); p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
...@@ -1879,6 +1905,43 @@ struct GridwiseGemm_xdl_cshuffle_v3 ...@@ -1879,6 +1905,43 @@ struct GridwiseGemm_xdl_cshuffle_v3
}); });
} }
} }
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run_2Lds(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
void* p_shared_0,
void* p_shared_1,
const Problem& problem)
{
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
Run_2Lds<decltype(a_grid_desc_ak0_m_ak1),
decltype(b_grid_desc_bk0_n_bk1),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
HasMainKBlockLoop,
CGlobalMemoryDataOperation,
TailNum>(p_a_grid,
p_b_grid,
p_c_grid,
p_shared_0,
p_shared_1,
problem,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock);
}
}; };
} // namespace ck } // namespace ck
include/ck_tile/core.hpp
View file @ 4947639c
...@@ -8,6 +8,7 @@ ...@@ -8,6 +8,7 @@
#include "ck_tile/core/algorithm/space_filling_curve.hpp" #include "ck_tile/core/algorithm/space_filling_curve.hpp"
#include "ck_tile/core/arch/amd_buffer_addressing.hpp" #include "ck_tile/core/arch/amd_buffer_addressing.hpp"
#include "ck_tile/core/arch/arch.hpp" #include "ck_tile/core/arch/arch.hpp"
#include "ck_tile/core/arch/generic_memory_space_atomic.hpp"
#include "ck_tile/core/arch/utility.hpp" #include "ck_tile/core/arch/utility.hpp"
#include "ck_tile/core/config.hpp" #include "ck_tile/core/config.hpp"
#include "ck_tile/core/container/array.hpp" #include "ck_tile/core/container/array.hpp"
...@@ -47,10 +48,12 @@ ...@@ -47,10 +48,12 @@
#include "ck_tile/core/tensor/tile_distribution_encoding.hpp" #include "ck_tile/core/tensor/tile_distribution_encoding.hpp"
#include "ck_tile/core/tensor/tile_elementwise.hpp" #include "ck_tile/core/tensor/tile_elementwise.hpp"
#include "ck_tile/core/tensor/tile_window.hpp" #include "ck_tile/core/tensor/tile_window.hpp"
#include "ck_tile/core/tensor/update_tile.hpp"
#include "ck_tile/core/utility/bit_cast.hpp" #include "ck_tile/core/utility/bit_cast.hpp"
#include "ck_tile/core/utility/functional.hpp" #include "ck_tile/core/utility/functional.hpp"
#include "ck_tile/core/utility/ignore.hpp" #include "ck_tile/core/utility/ignore.hpp"
#include "ck_tile/core/utility/magic_div.hpp" #include "ck_tile/core/utility/magic_div.hpp"
#include "ck_tile/core/utility/philox_rand.hpp"
#include "ck_tile/core/utility/random.hpp" #include "ck_tile/core/utility/random.hpp"
#include "ck_tile/core/utility/to_sequence.hpp" #include "ck_tile/core/utility/to_sequence.hpp"
#include "ck_tile/core/utility/transpose_vectors.hpp" #include "ck_tile/core/utility/transpose_vectors.hpp"
......
include/ck_tile/core/arch/generic_memory_space_atomic.hpp 0 → 100644
View file @ 4947639c
This diff is collapsed.
include/ck_tile/core/config.hpp
View file @ 4947639c
...@@ -171,3 +171,7 @@ ...@@ -171,3 +171,7 @@
#ifndef CK_TILE_FMHA_FWD_FAST_EXP2 #ifndef CK_TILE_FMHA_FWD_FAST_EXP2
#define CK_TILE_FMHA_FWD_FAST_EXP2 0 #define CK_TILE_FMHA_FWD_FAST_EXP2 0
#endif #endif
#ifndef CK_TILE_BUFFER_LOAD_RAW_BF16_WA
#define CK_TILE_BUFFER_LOAD_RAW_BF16_WA 1
#endif
include/ck_tile/core/numeric/vector_type.hpp
View file @ 4947639c
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -144,6 +144,15 @@ using int8x16_t = int8_t __attribute((ext_vector_type(16))); ...@@ -144,6 +144,15 @@ using int8x16_t = int8_t __attribute((ext_vector_type(16)));
using int8x32_t = int8_t __attribute((ext_vector_type(32))); using int8x32_t = int8_t __attribute((ext_vector_type(32)));
using int8x64_t = int8_t __attribute((ext_vector_type(64))); using int8x64_t = int8_t __attribute((ext_vector_type(64)));
// ui8
// using uint8_t
using uint8x2_t = uint8_t __attribute((ext_vector_type(2)));
using uint8x4_t = uint8_t __attribute((ext_vector_type(4)));
using uint8x8_t = uint8_t __attribute((ext_vector_type(8)));
using uint8x16_t = uint8_t __attribute((ext_vector_type(16)));
using uint8x32_t = uint8_t __attribute((ext_vector_type(32)));
using uint8x64_t = uint8_t __attribute((ext_vector_type(64)));
#if CK_TILE_USE_CUSTOM_DATA_TYPE #if CK_TILE_USE_CUSTOM_DATA_TYPE
// f8 // f8
// using fp8_t // using fp8_t
......
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