Skip to content

GitLab

Commit d64030ed authored by aska-0096's avatar aska-0096
Browse files

revert blockwisegemm modification

parent 1c7f994f
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 296 additions and 2494 deletions
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1.hpp
View file @ d64030ed
...@@ -155,158 +155,6 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave, ...@@ -155,158 +155,6 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
return TailNumber::Full; return TailNumber::Full;
} }
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
// -------------------------------------------------------------------------------------------
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
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>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
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>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
}
template <bool HasMainLoop, template <bool HasMainLoop,
TailNumber TailNum, TailNumber TailNum,
typename AGridDesc, typename AGridDesc,
...@@ -480,6 +328,7 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave, ...@@ -480,6 +328,7 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
} }
} }
protected:
using Base::a_thread_copy_; using Base::a_thread_copy_;
using Base::a_thread_desc_; using Base::a_thread_desc_;
using Base::b_thread_copy_; using Base::b_thread_copy_;
...@@ -607,191 +456,6 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave, ...@@ -607,191 +456,6 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave,
return TailNumber::Full; return TailNumber::Full;
} }
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
// -------------------------------------------------------------------------------------------
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
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>{}),
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 and there's no
// observable negative impact. The desired effect is waves in a workgroup
// executing MAC in sync. This avoids some out-of-sync waves hijacking MAC
// resource from other workgroups and reducing the chance of latency hiding by
// waiting for the rest of the workgroup at the eventual sync point.
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
// The block_sync_lds() here performs double duty:
// A) safeguard against data hazard because barrier from
// blockwise_gemm is moved here B) reduce VMEM FIFO congestion by
// applying small delays to different wavefronts It is performed
// near the end of MAC cluster to minimize lgkmcnt penalty
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
// block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
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>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
}
template <bool HasMainLoop, template <bool HasMainLoop,
TailNumber TailNum, TailNumber TailNum,
typename AGridDesc, typename AGridDesc,
...@@ -1023,6 +687,7 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave, ...@@ -1023,6 +687,7 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave,
} }
} }
protected:
// K->M loopover // K->M loopover
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor( static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}), make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}),
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp
View file @ d64030ed
...@@ -262,227 +262,6 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave, ...@@ -262,227 +262,6 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
}); });
} }
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
__builtin_amdgcn_sched_barrier(0);
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefetch 1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
i += 1;
} while(i < (num_loop - 2));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf_tail);
});
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf_tail);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
}
}
template <bool HasMainLoop, template <bool HasMainLoop,
TailNumber TailNum, TailNumber TailNum,
typename AGridDesc, typename AGridDesc,
...@@ -635,69 +414,11 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave, ...@@ -635,69 +414,11 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
__builtin_amdgcn_sched_barrier(0); __builtin_amdgcn_sched_barrier(0);
i += 1; i += 1;
} while(i < (num_loop - 2)); } while(i < (num_loop - 1));
} }
// tail // tail
if constexpr(TailNum == TailNumber::Full) if constexpr(TailNum == TailNumber::Full)
{ {
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
static_for<0, KRepeat, 1>{}([&](auto k0) { static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, NRepeat, 1>{}([&](auto n0) {
...@@ -731,6 +452,7 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave, ...@@ -731,6 +452,7 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
} }
} }
protected:
using Base::a_thread_copy_; using Base::a_thread_copy_;
using Base::a_thread_desc_; using Base::a_thread_desc_;
using Base::b_thread_copy_; using Base::b_thread_copy_;
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp
View file @ d64030ed
...@@ -234,316 +234,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave, ...@@ -234,316 +234,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
__builtin_amdgcn_sched_barrier(0); __builtin_amdgcn_sched_barrier(0);
} }
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0));
// Local prefetch 1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(I0),
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>{}),
b_block_buf.At(I0),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(I0));
});
});
});
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 2
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1));
// Global prefetch 3
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
// This hot loop has two legacy loopover, to implement the double local buffer strategy
do
{
auto LoopFunc = [&](auto lds_read_buf,
auto lds_read_reg_buf,
auto lds_write_buf,
auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
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>{}),
b_block_buf.At(lds_read_buf),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf));
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf]
[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf]
[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
};
LoopFunc(I1, I1, I0, I0);
LoopFunc(I0, I0, I1, I1);
i += HotloopUnroll;
} while(i < (num_loop - PrefetchStages));
}
auto ReadWriteCompFunc = [&](auto lds_read_buf,
auto lds_read_reg_buf,
auto lds_write_buf,
auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
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>{}),
b_block_buf.At(lds_read_buf),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf));
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
};
auto ReadCompFunc = [&](auto lds_read_buf, auto lds_read_reg_buf, auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
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>{}),
b_block_buf.At(lds_read_buf),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
};
auto CompFunc = [&](auto mfma_reg_buf) {
a_thread_buf_tail = a_thread_bufs[mfma_reg_buf];
b_thread_buf_tail = b_thread_bufs[mfma_reg_buf];
};
// tail
if constexpr(TailNum == TailNumber::Odd)
{
ReadWriteCompFunc(I1, I1, I0, I0);
ReadCompFunc(I0, I0, I1);
CompFunc(I0);
}
else if constexpr(TailNum == TailNumber::Even)
{
ReadCompFunc(I1, I1, I0);
CompFunc(I1);
}
}
template <bool HasMainLoop, template <bool HasMainLoop,
TailNumber TailNum, TailNumber TailNum,
typename AGridDesc, typename AGridDesc,
...@@ -873,6 +563,7 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave, ...@@ -873,6 +563,7 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
} }
} }
protected:
using Base::a_thread_copy_; using Base::a_thread_copy_;
using Base::a_thread_desc_; using Base::a_thread_desc_;
using Base::b_thread_copy_; using Base::b_thread_copy_;
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp
View file @ d64030ed
...@@ -316,270 +316,6 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave, ...@@ -316,270 +316,6 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
__builtin_amdgcn_sched_barrier(0); __builtin_amdgcn_sched_barrier(0);
} }
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_loop.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_loop.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Global prefetch 3
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefetch 1
block_sync_lds();
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_loop.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, I0),
a_block_buf,
a_thread_desc_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_loop.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, I0),
b_block_buf,
b_thread_desc_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
auto LoopFunc = [&](auto vmem_buf) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KRepeat, 1>{}([&](auto k0) {
if constexpr(k0 == (KRepeat - 1))
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, vmem_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, vmem_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
}
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_loop.CalculateOffset(
make_tuple(m0, I0, I0, ik))>{}];
});
static_for<0, KPack, 1>{}([&](auto ik) {
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_loop.CalculateOffset(
make_tuple(n0, I0, I0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
a_thread_copy_loop.Run(
a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
a_block_buf,
a_thread_desc_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_loop.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
b_block_buf,
b_thread_desc_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
};
LoopFunc(I0);
LoopFunc(I1);
i += HotloopUnroll;
} while(i < (num_loop - PrefetchStages));
}
// tail
auto ReadWriteCompFunc = [&](auto vmem_buf) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KRepeat, 1>{}([&](auto k0) {
if constexpr(k0 == (KRepeat - 1))
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
block_sync_lds();
}
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_loop.CalculateOffset(
make_tuple(m0, I0, I0, ik))>{}];
});
static_for<0, KPack, 1>{}([&](auto ik) {
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_loop.CalculateOffset(
make_tuple(n0, I0, I0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
a_thread_copy_loop.Run(
a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
a_block_buf,
a_thread_desc_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_loop.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
b_block_buf,
b_thread_desc_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
};
auto ReadCompFunc = [&]() {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf_tail);
});
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf_tail);
});
});
HotLoopScheduler();
};
if constexpr(TailNum == TailNumber::Odd)
{
ReadWriteCompFunc(I0);
ReadWriteCompFunc(I1);
ReadCompFunc();
}
else if constexpr(TailNum == TailNumber::Even)
{
ReadWriteCompFunc(I0);
ReadCompFunc();
}
}
template <bool HasMainLoop, template <bool HasMainLoop,
TailNumber TailNum, TailNumber TailNum,
typename AGridDesc, typename AGridDesc,
...@@ -891,18 +627,19 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave, ...@@ -891,18 +627,19 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
} }
} }
protected:
// A[MRepeat, I1, I1, KPack] // A[MRepeat, I1, I1, KPack]
static constexpr auto a_thread_desc_loop = static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I1, I1, Number<KPack>{})); make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I1, I1, Number<KPack>{}));
// B[NRepeat, N1, N2, KPack] // B[NRepeat, N1, N2, KPack]
static constexpr auto b_thread_desc_loop = static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<NRepeat>{}, I1, I1, Number<KPack>{})); make_naive_tensor_descriptor_packed(make_tuple(Number<NRepeat>{}, I1, I1, Number<KPack>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType, using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
ComputeDataType, ComputeDataType,
decltype(a_block_desc_m0_m1_m2_k), decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_loop), decltype(a_thread_desc_),
Sequence<1, 1, 1, KPack>, Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>, Sequence<0, 1, 2, 3>,
3, 3,
...@@ -912,19 +649,15 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave, ...@@ -912,19 +649,15 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType, using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
ComputeDataType, ComputeDataType,
decltype(b_block_desc_n0_n1_n2_k), decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_loop), decltype(b_thread_desc_),
Sequence<1, 1, 1, KPack>, Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>, Sequence<0, 1, 2, 3>,
3, 3,
B_K1, B_K1,
B_K1>; B_K1>;
AThreadCopy a_thread_copy_loop{Base::CalculateAThreadOriginDataIndex()}; AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_loop{Base::CalculateBThreadOriginDataIndex()}; BThreadCopy b_thread_copy_{Base::CalculateBThreadOriginDataIndex()};
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
using Base::b_thread_desc_;
using Base::c_thread_desc_; using Base::c_thread_desc_;
}; };
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp
View file @ d64030ed
...@@ -1220,38 +1220,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1220,38 +1220,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
return c_grid_desc_mblock_mperblock_nblock_nperblock; return c_grid_desc_mblock_mperblock_nblock_nperblock;
} }
__device__ static constexpr auto EpilogueScheduler()
{
constexpr auto epilogue_tile = MPerBlock * NPerBlock * CShuffleMXdlPerWavePerShuffle *
CShuffleNXdlPerWavePerShuffle / (MXdlPerWave * NXdlPerWave);
constexpr auto num_mfma_inst = BlockwiseGemmPipe::HotLoopInstList::C_MFMA_Inst_Num *
CShuffleMXdlPerWavePerShuffle *
CShuffleNXdlPerWavePerShuffle / (MXdlPerWave * NXdlPerWave);
constexpr auto num_ds_write_inst =
epilogue_tile / BlockSize; // DefaultMFMA, per-element write
constexpr auto num_ds_read_inst =
epilogue_tile / BlockSize / CShuffleBlockTransferScalarPerVector_NPerBlock;
constexpr auto num_buffer_store_inst = num_ds_read_inst;
// MFMA:ds_write=1:2
constexpr auto num_ds_write_issue = num_ds_write_inst / 2;
constexpr auto num_mfma_block_sync = (num_mfma_inst - num_ds_write_issue) / 2;
constexpr auto mfma_ds_write_rate = MXdlPerWave == 16 ? 2 : 4;
// Hide ds_write issue latency
static_for<0, num_ds_write_issue, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, mfma_ds_write_rate, 0); // DS write
});
// Hide block_sync + ds_read latency
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_block_sync, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, num_ds_read_inst, 0); // DS read
// Hide block_sync latency
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_block_sync, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x040, num_buffer_store_inst, 0); // VMEM write
}
// return block_id to C matrix tile idx (m0, n0) mapping // return block_id to C matrix tile idx (m0, n0) mapping
// if arch = gfx942 // if arch = gfx942
using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>; using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
...@@ -1429,15 +1397,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1429,15 +1397,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{}; auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer(); auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
constexpr auto a_thread_desc = blockwise_gemm_pipeline.a_thread_desc_;
constexpr auto b_thread_desc = blockwise_gemm_pipeline.b_thread_desc_;
constexpr auto c_thread_desc = blockwise_gemm_pipeline.c_thread_desc_;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
b_thread_desc.GetElementSpaceSize());
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane( const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) / (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock); KPerBlock);
...@@ -1455,16 +1414,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1455,16 +1414,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
b_block_buf, b_block_buf,
b_block_slice_copy_step, b_block_slice_copy_step,
c_thread_buf, c_thread_buf,
a_thread_buf,
b_thread_buf,
num_k_block_main_loop); num_k_block_main_loop);
// shuffle C and write out // shuffle C and write out
{ {
// Last block MFMA
auto xdlops_gemm = blockwise_gemm_pipeline.xdlops_gemm;
constexpr auto KRepeat = blockwise_gemm_pipeline.KRepeat;
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 && static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0, NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!"); "wrong!");
...@@ -1624,9 +1577,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1624,9 +1577,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock; CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation; const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
// C: LDS -> VGPR
// D: Global -> VGPR
// E: =Epilogue(C, D), VGPR -> Global
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3< auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3<
ThisThreadBlock, ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})), decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
...@@ -1685,84 +1635,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1685,84 +1635,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!"); static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
constexpr auto KPerInnerLoop = blockwise_gemm_pipeline.KPerInnerLoop;
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
});
__builtin_amdgcn_sched_barrier(0);
static_for<0, num_access, 1>{}([&](auto access_id) { static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS // make sure it's safe to write to LDS
block_sync_lds(); block_sync_lds();
if constexpr(access_id < num_access - 1)
{
constexpr auto shuffle_m0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<0>{}];
constexpr auto shuffle_n0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<1>{}];
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(shuffle_n0 + n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = c_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, shuffle_n0 + n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
});
}
// each thread write its data from VGPR to LDS // each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2, c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id), sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
...@@ -1796,8 +1672,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1796,8 +1672,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
tie(e_grid_desc_mblock_mperblock_nblock_nperblock), tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0, I0,
cde_lds_and_global_step); cde_lds_and_global_step);
// EpilogueScheduler();
} }
}); });
} }
...@@ -1990,15 +1864,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -1990,15 +1864,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{}; auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer(); auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
constexpr auto a_thread_desc = blockwise_gemm_pipeline.a_thread_desc_;
constexpr auto b_thread_desc = blockwise_gemm_pipeline.b_thread_desc_;
constexpr auto c_thread_desc = blockwise_gemm_pipeline.c_thread_desc_;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
b_thread_desc.GetElementSpaceSize());
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane( const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) / (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock); KPerBlock);
...@@ -2016,16 +1881,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -2016,16 +1881,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
b_block_bufs, b_block_bufs,
b_block_slice_copy_step, b_block_slice_copy_step,
c_thread_buf, c_thread_buf,
a_thread_buf,
b_thread_buf,
num_k_block_main_loop); num_k_block_main_loop);
// shuffle C and write out // shuffle C and write out
{ {
// Last block MFMA
auto xdlops_gemm = blockwise_gemm_pipeline.xdlops_gemm;
constexpr auto KRepeat = blockwise_gemm_pipeline.KRepeat;
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 && static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0, NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!"); "wrong!");
...@@ -2243,84 +2102,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -2243,84 +2102,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!"); static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
constexpr auto KPerInnerLoop = blockwise_gemm_pipeline.KPerInnerLoop;
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
});
__builtin_amdgcn_sched_barrier(0);
static_for<0, num_access, 1>{}([&](auto access_id) { static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS // make sure it's safe to write to LDS
block_sync_lds(); block_sync_lds();
if constexpr(access_id < num_access - 1)
{
constexpr auto shuffle_m0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<0>{}];
constexpr auto shuffle_n0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<1>{}];
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(shuffle_n0 + n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = c_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, shuffle_n0 + n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
});
}
// each thread write its data from VGPR to LDS // each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2, c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id), sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
...@@ -2354,8 +2139,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3 ...@@ -2354,8 +2139,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
tie(e_grid_desc_mblock_mperblock_nblock_nperblock), tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0, I0,
cde_lds_and_global_step); cde_lds_and_global_step);
// EpilogueScheduler();
} }
}); });
} }
......
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