squall/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity
c008b4aea8
cutlass/include/cute/atom/copy_traits_sm90_tma.hpp
Pradeep Ramani c008b4aea8
CUTLASS 3.3.0 (#1167) 
* Release 3.3.0

Adds support for mixed precision GEMMs On Hopper and Ampere
Adds support for < 16B aligned GEMMs on Hopper
Enhancements to EVT
Enhancements to Python interface
Enhancements to Sub-byte type handling in CuTe
Several other bug-fixes and performance improvements.

* minor doc update
2023-11-02 11:09:05 -04:00

1057 lines
43 KiB
C++
Raw Blame History

/***************************************************************************************************
* Copyright (c) 2023 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* 3. Neither the name of the copyright holder nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
#pragma once
#if !defined(__CUDACC_RTC__)
#include <cuda.h>
#endif
#include <cute/atom/copy_traits_sm90_tma_swizzle.hpp>
#include <cute/atom/copy_traits.hpp>
#include <cute/atom/copy_atom.hpp>
#include <cute/numeric/integral_ratio.hpp>
namespace cute
{
template <class GmemStrides_, class TmaGBasis_, class TmaSwizzle_>
struct AuxTmaParams {
using GmemStrides = GmemStrides_;
GmemStrides g_stride_;
};
//////////////////////////////////////////////////////////////////////////////
///////////////////////////// TMA_LOAD ///////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////
struct SM90_TMA_LOAD_OP : SM90_TMA_LOAD {};
// The executable SM90_TMA_LOAD with tma_desc and tma_mbar
template <class NumBitsPerTMA>
struct Copy_Traits<SM90_TMA_LOAD_OP, NumBitsPerTMA>
{
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
// SM90_TMA_LOAD arguments
TmaDescriptor const& tma_desc_;
uint64_t& tma_load_mbar_;
template <class Coord, int... Is>
CUTE_HOST_DEVICE constexpr
void
copy_unpack_(void const* const dst_ptr,
Coord const& src_coord, seq<Is...>) const
{
#if 0
auto [c0,c1,c2,c3,c4] = append<5>(src_coord, 0);
printf("THR (%d,%d,%d) BLK (%d,%d,%d) TMACRD (%d,%d,%d,%d,%d) SMEMADDR (%p)\n",
threadIdx.x, threadIdx.y, threadIdx.z,
blockIdx.x, blockIdx.y, blockIdx.z,
int32_t(c0), int32_t(c1), int32_t(c2), int32_t(c3), int32_t(c4), dst_ptr);
#endif
SM90_TMA_LOAD::copy(&tma_desc_, tma_load_mbar_,
dst_ptr, get<Is>(src_coord)...);
}
// This is the copy_unpack dispatch for this Copy_Traits
// Src needs to be a gmem tensor with TmaCoordIterator .data()
// Dst needs to be a smem tensor
template <class TS, class SLayout,
class TD, class DLayout>
CUTE_HOST_DEVICE friend constexpr
void
copy_unpack(Copy_Traits const& traits,
Tensor<TS,SLayout> const& src,
Tensor<TD,DLayout> & dst)
{
static_assert(is_smem<TD>::value, "Expected smem dst for SM90_TMA_LOAD");
traits.copy_unpack_(cute::raw_pointer_cast(dst.data()), src.data().coord_, tuple_seq<decltype(src.data().coord_)>{});
}
};
// The non-executable SM90_TMA_LOAD with tma_desc and no tma_mbar
// Use .with(tma_mbar) to construct an executable version
template <class NumBitsPerTMA, class AuxParams_>
struct Copy_Traits<SM90_TMA_LOAD, NumBitsPerTMA, AuxParams_>
{
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
// SM90_TMA_LOAD arguments
TmaDescriptor tma_desc_;
using AuxParams = AuxParams_;
AuxParams aux_params_;
// Return TmaDescriptor/TensorMap
CUTE_HOST_DEVICE constexpr
TmaDescriptor const*
get_tma_descriptor() const {
return &tma_desc_;
}
// Construct an executable SM90_TMA_LOAD with tma_mbar
CUTE_HOST_DEVICE constexpr
Copy_Traits<SM90_TMA_LOAD_OP, NumBitsPerTMA>
with(uint64_t& tma_mbar, uint16_t const& multicast_mask = 0) const {
// We accept multicast_mask here to keep the API for both atoms consistent
// assert(multicast_mask == 0);
(void) multicast_mask;
return {tma_desc_, tma_mbar};
}
// Generate the TMA coord tensor
template <class GShape>
CUTE_HOST_DEVICE constexpr
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
// Don't try to execute a copy with SM90_TMA_LOAD before calling .with()
template <class TS, class SLayout,
class TD, class DLayout>
CUTE_HOST_DEVICE friend constexpr void
copy_unpack(Copy_Traits const& traits,
Tensor<TS,SLayout> const& src,
Tensor<TD,DLayout> & dst) = delete;
};
//////////////////////////////////////////////////////////////////////////////
///////////////////////////// TMA_LOAD_MULTICAST /////////////////////////////
//////////////////////////////////////////////////////////////////////////////
struct SM90_TMA_LOAD_MULTICAST_OP : SM90_TMA_LOAD_MULTICAST {};
template <class NumBitsPerTMA>
struct Copy_Traits<SM90_TMA_LOAD_MULTICAST_OP, NumBitsPerTMA>
{
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
// SM90_TMA_LOAD_MULTICAST arguments
TmaDescriptor const& tma_desc_;
uint64_t& tma_load_mbar_;
uint16_t const& multicast_mask_;
template <class Coord, int... Is>
CUTE_HOST_DEVICE constexpr
void
copy_unpack_(void const* const dst_ptr,
Coord const& src_coord, seq<Is...>) const
{
#if 0
auto [c0,c1,c2,c3,c4] = append<5>(src_coord, 0);
printf("THR (%d,%d,%d) BLK (%d,%d,%d) TMACRD (%d,%d,%d,%d,%d) SMEMADDR (%p)\n",
threadIdx.x, threadIdx.y, threadIdx.z,
blockIdx.x, blockIdx.y, blockIdx.z,
int32_t(c0), int32_t(c1), int32_t(c2), int32_t(c3), int32_t(c4), dst_ptr);
#endif
SM90_TMA_LOAD_MULTICAST::copy(&tma_desc_, tma_load_mbar_, multicast_mask_,
dst_ptr, get<Is>(src_coord)...);
}
template <class TS, class SLayout,
class TD, class DLayout>
CUTE_HOST_DEVICE friend constexpr
void
copy_unpack(Copy_Traits const& traits,
Tensor<TS,SLayout> const& src,
Tensor<TD,DLayout> & dst)
{
static_assert(is_smem<TD>::value, "Expected smem dst for SM90_TMA_LOAD_MULTICAST");
traits.copy_unpack_(cute::raw_pointer_cast(dst.data()), src.data().coord_, tuple_seq<decltype(src.data().coord_)>{});
}
};
template <class NumBitsPerTMA, class AuxParams_>
struct Copy_Traits<SM90_TMA_LOAD_MULTICAST, NumBitsPerTMA, AuxParams_>
{
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
// SM90_TMA_LOAD_MULTICAST arguments
TmaDescriptor tma_desc_;
using AuxParams = AuxParams_;
AuxParams aux_params_;
// Return TmaDescriptor/TensorMap
CUTE_HOST_DEVICE constexpr
TmaDescriptor const*
get_tma_descriptor() const {
return &tma_desc_;
}
// Construct an executable SM90_TMA_LOAD_MULTICAST with tma_mbar
CUTE_HOST_DEVICE constexpr
Copy_Traits<SM90_TMA_LOAD_MULTICAST_OP, NumBitsPerTMA>
with(uint64_t& tma_load_mbar, uint16_t const& multicast_mask) const {
return {tma_desc_, tma_load_mbar, multicast_mask};
}
// Generate the TMA coord tensor
template <class GShape>
CUTE_HOST_DEVICE constexpr
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
// Don't try to execute a copy with SM90_TMA_LOAD_MULTICAST before calling .with()
template <class TS, class SLayout,
class TD, class DLayout>
CUTE_HOST_DEVICE friend constexpr void
copy_unpack(Copy_Traits const& traits,
Tensor<TS,SLayout> const& src,
Tensor<TD,DLayout> & dst) = delete;
};
//////////////////////////////////////////////////////////////////////////////
///////////////////////////// TMA_STORE //////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////
// The executable SM90_TMA_STORE with tma_desc
template <class NumBitsPerTMA, class AuxParams_>
struct Copy_Traits<SM90_TMA_STORE, NumBitsPerTMA, AuxParams_>
{
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
// SM90_TMA_STORE arguments
TmaDescriptor tma_desc_;
using AuxParams = AuxParams_;
AuxParams aux_params_;
// Return TmaDescriptor/TensorMap
CUTE_HOST_DEVICE constexpr
TmaDescriptor const*
get_tma_descriptor() const {
return &tma_desc_;
}
// Generate the TMA coord tensor
template <class GShape>
CUTE_HOST_DEVICE constexpr
auto
get_tma_tensor(GShape const& g_shape) const {
static_assert(is_congruent<decltype(g_shape), decltype(aux_params_.g_stride_)>::value);
return make_counting_tensor(make_layout(g_shape, aux_params_.g_stride_));
}
template <class Coord, int... Is>
CUTE_HOST_DEVICE constexpr
void
copy_unpack_(void const* const src_ptr,
Coord const& dst_coord, seq<Is...>) const
{
#if 0
auto [c0,c1,c2,c3,c4] = append<5>(dst_coord, 0);
printf("THR (%d,%d,%d) BLK (%d,%d,%d) TMACRD (%d,%d,%d,%d,%d) SMEMADDR (%p)\n",
threadIdx.x, threadIdx.y, threadIdx.z,
blockIdx.x, blockIdx.y, blockIdx.z,
int32_t(c0), int32_t(c1), int32_t(c2), int32_t(c3), int32_t(c4), src_ptr);
#endif
SM90_TMA_STORE::copy(&tma_desc_,
src_ptr, get<Is>(dst_coord)...);
}
// This is the copy_unpack dispatch for this Copy_Traits
// Src needs to be a smem tensor
// Dst needs to be a gmem tensor with TmaCoordIterator .data()
template <class TS, class SLayout,
class TD, class DLayout>
CUTE_HOST_DEVICE friend constexpr
void
copy_unpack(Copy_Traits const& traits,
Tensor<TS,SLayout> const& src,
Tensor<TD,DLayout> & dst)
{
static_assert(is_smem<TS>::value, "Expected smem src for SM90_TMA_STORE");
//static_assert(is_gmem<TD>::value, "Expected gmem dst for SM90_TMA_STORE"); // TMA spoofed src tensor
traits.copy_unpack_(cute::raw_pointer_cast(src.data()), dst.data().coord_, tuple_seq<decltype(dst.data().coord_)>{});
}
};
//////////////////////////////////////////////////////////////////////////////
///////////////////////////// BULK COPY //////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////
template <class NumBitsPerTMA, class... OpArgs>
struct Copy_Traits<SM90_BULK_COPY_G2S, NumBitsPerTMA, OpArgs...>
{
static_assert(int32_t(NumBitsPerTMA::value / 8) % 16 == 0,
"Bulk Copy requires copy vector size align to 16B.");
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
// SM90_BULK_COPY_G2S arguments
// 0: uint64_t* bulk_load_memory_barrier
cute::tuple<OpArgs...> bulk_load_mbar_;
template <class TS, class SLayout,
class TD, class DLayout>
CUTE_HOST_DEVICE friend constexpr
void
copy_unpack(Copy_Traits const& traits,
Tensor<TS,SLayout> const& src,
Tensor<TD,DLayout> & dst)
{
static_assert(is_same<cute::tuple<OpArgs...>, cute::tuple<uint64_t*>>::value,
"Extra arguments not set. Set .with() before use.");
static_assert(is_gmem<TS>::value, "Expected gmem src for SM90_BULK_COPY_G2S");
static_assert(is_smem<TD>::value, "Expected smem dst for SM90_BULK_COPY_G2S");
SM90_BULK_COPY_G2S::copy(raw_pointer_cast(src.data()), *get<0>(traits.bulk_load_mbar_),
raw_pointer_cast(dst.data()), int32_t(NumBitsPerTMA::value / 8));
}
// Record the memory barrier for the instruction
CUTE_HOST_DEVICE constexpr
Copy_Traits<SM90_BULK_COPY_G2S, NumBitsPerTMA, uint64_t*>
with(uint64_t& bulk_mbar) const {
return {{&bulk_mbar}};
}
};
template <class NumBitsPerTMA>
struct Copy_Traits<SM90_BULK_COPY_S2G, NumBitsPerTMA>
{
static_assert(int32_t(NumBitsPerTMA::value / 8) % 16 == 0,
"Bulk Copy requires copy vector size align to 16B.");
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,NumBitsPerTMA>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
template <class TS, class SLayout,
class TD, class DLayout>
CUTE_HOST_DEVICE friend constexpr
void
copy_unpack(Copy_Traits const& traits,
Tensor<TS,SLayout> const& src,
Tensor<TD,DLayout> & dst)
{
static_assert(is_smem<TS>::value, "Expected smem src for SM90_BULK_COPY_S2G");
static_assert(is_gmem<TD>::value, "Expected gmem dst for SM90_BULK_COPY_S2G");
SM90_BULK_COPY_S2G::copy(raw_pointer_cast(src.data()), raw_pointer_cast(dst.data()), int32_t(NumBitsPerTMA::value / 8));
}
};
//
// Placeholder for the bulk copy algorithm's default, auto-vectorizing behavior
//
template <class... OpArgs>
struct Copy_Traits<SM90_BULK_COPY_AUTO, OpArgs...>
{
// Logical thread id to thread idx (one-thread)
using ThrID = Layout<_1>;
// Map from (src-thr,src-val) to bit
using SrcLayout = Layout<Shape<_1,_1>, Stride<_0,_0>>;
// Map from (dst-thr,dst-val) to bit
using DstLayout = Layout<Shape<_1,_1>, Stride<_0,_0>>;
// Reference map from (thr,val) to bit
using RefLayout = SrcLayout;
// SM90_UBULK_COPY arguments
// 0: uint64_t* bulk_load_memory_barrier [if this is a BULK_LOAD_G2S]
cute::tuple<OpArgs...> opargs_;
// Record the memory barrier for the instruction
CUTE_HOST_DEVICE constexpr
Copy_Traits<SM90_BULK_COPY_AUTO, uint64_t*>
with(uint64_t& bulk_mbar) const {
return {{&bulk_mbar}};
}
};
//
// MAKE_TMA_COPY and related
//
namespace detail {
// Custom version of coalesce that greedily combines modes only up to size-256
// Look at each element and the back of the stack (in order of priority)
// back(NewLayout) get<I>(OldLayout)
// s0:d0 _1:d1 => continue
// _1:d0 s1:d1 => replace_back s1:d1
// s0:d0 s1:s0*d0 => replace_back s0*s1:d0 if s0*s1 <= 256
// s0:d0 s1:d1 => append s1:d1
//
// @pre OldShape and OldStride are flat
template <int I, class OldShape, class OldStride, class NewShape, class NewStride>
CUTE_HOST_DEVICE constexpr
auto
coalesce_256_impl(OldShape const& old_shape, OldStride const& old_stride,
NewShape const& new_shape, NewStride const& new_stride)
{
if constexpr (I == rank_v<OldShape>) {
// Base case, we're done
if constexpr (is_constant<1, NewShape>::value) {
return Layout<_1,_0>{};
} else {
return Layout<NewShape,NewStride>{new_shape,new_stride};
}
} else if constexpr (is_constant<1, decltype(get<I>(old_shape))>::value) {
// shape<I>(layout) == _1, skip it and continue
return coalesce_256_impl<I+1>(old_shape, old_stride, new_shape, new_stride);
} else if constexpr (is_constant<1, NewShape>::value) {
// Replace our shape-1 with anything (Can only happen on input new_shape/new_stride)
return coalesce_256_impl<I+1>(old_shape, old_stride, get<I>(old_shape), get<I>(old_stride));
} else if constexpr (is_constant<true, decltype(back(new_shape) * back(new_stride) == get<I>(old_stride) &&
get<I>(old_shape) * back(new_shape) <= Int<256>{})>::value) {
// Merge modes because the shapes and strides match and the merge is 256 or less
return coalesce_256_impl<I+1>(old_shape, old_stride,
replace_back(new_shape, get<I>(old_shape) * back(new_shape)),
new_stride);
} else {
// Can't replace or merge, so append a new mode
return coalesce_256_impl<I+1>(old_shape, old_stride,
append(new_shape, get<I>(old_shape)),
append(new_stride, get<I>(old_stride)));
}
CUTE_GCC_UNREACHABLE;
}
// Combine all the modes that are possible to combine
// Does not respect the profile of the layout, but does preserve total size
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
coalesce_256(Layout<Shape,Stride> const& layout)
{
auto flat_shape = flatten(layout.shape());
auto flat_stride = flatten(layout.stride());
return coalesce_256_impl<1>(flat_shape, flat_stride, get<0>(flat_shape), get<0>(flat_stride));
}
template <class Engine, class Layout>
CUTE_HOST_DEVICE constexpr
auto
coalesce_256(Tensor<Engine,Layout> const& tensor)
{
return make_tensor(tensor.data(), coalesce_256(tensor.layout()));
}
// Use a smem_inv_h to read through the GMEM tensor
// and construct a TMA Descriptor for the resulting instruction
// At the same time, construct the Tma Tensor's Stride to generate
// the TMA coordinates that the instruction consumes.
//
template <class TmaInternalType,
class GEngine, class GLayout,
class SShape, class SStride,
int B, int M, int S>
CUTE_HOST_RTC
auto
make_tma_copy_desc(Tensor<GEngine,GLayout> const& gtensor, // The original GMEM Tensor
Layout<SShape,SStride> const& smem_inv_h, // smem_idx to hier gmode
Swizzle<B,M,S> const& swizzle) // Swizzle fn on smem_idx
{
// The smem vector is the same units as gtensor, so compose first and then recast
// tma_val_idx:gmem_strides
Tensor tile_gstride = recast<TmaInternalType>(gtensor.compose(smem_inv_h));
// Coalesce modes up to size-256 (the maximum TMA box extent in units of TmaInternalType)
// tma_box_shape:gmem_strides
Tensor tma_gstride = coalesce_256(tile_gstride);
// Perform the tiling to the gmem vector again, but with indirections to the gtensor modes
auto gbasis = make_identity_layout(shape(gtensor));
auto tile_gbasis_tmp = gbasis.compose(smem_inv_h);
// Instead of the recast (gbasis doesn't have type info), replace the shape with the already-recasted shape
// tma_box_shape:gmem_mode
auto tile_gbasis = make_layout(shape(tile_gstride), stride(tile_gbasis_tmp));
// Recast the original tensor for shape inspections
auto gtensor_T = recast<TmaInternalType>(gtensor);
// Find missing bases that don't appear in tile_gbasis
// NOTE This is essentially ArithmeticTuple complement...
// NOTE in pursuit of implementing an ArithmeticTuple logical_divide for smem_inv_h
auto tile_gbasis_remaining_stride = filter_tuple(flatten(shape (gtensor_T)), flatten(stride(gtensor_T)),
flatten(stride(gbasis)),
[&](auto s, auto d, auto e)
{
if constexpr (is_constant<1, decltype(s)>::value || is_constant<0, decltype(d)>::value) {
return cute::tuple<>{}; // If size-1 or stride-0, then don't append
} else {
using E = decltype(e);
auto has_e = any_of(stride(tile_gbasis), [] (auto tb) { return tb == E{}; });
if constexpr (decltype(has_e)::value) {
return cute::tuple<>{}; // If d was found, then don't append
} else {
return cute::tuple<E>(e); // Else, this is missing so append
}
}
});
auto tile_gbasis_remaining_rank = rank(tile_gbasis_remaining_stride);
// "Coalesce" the tile basis into a compatible shape with the tma
auto tma_gbasis_tile = tile_gbasis.compose(make_layout(wrap(shape(tma_gstride))));
// Append the remaining basis modes that contribute to the TMA with size-1
auto tma_gbasis_full = make_layout(tuple_cat(wrap( shape(tma_gbasis_tile)), wrap(repeat<tile_gbasis_remaining_rank>(Int<1>{}))),
tuple_cat(wrap(stride(tma_gbasis_tile)), wrap(tile_gbasis_remaining_stride)));
// Group the trailing modes to make this max rank-5 -- TMA rank limitation
// tma_box_shape:gmem_mode
auto tma_gbasis = group<cute::min(rank(tma_gbasis_full),4),-1>(tma_gbasis_full);
#if 0
print("smem_inv_h : "); print(smem_inv_h); print("\n");
print("gtensor : "); print(gtensor); print("\n");
print("tile_gstride : "); print(tile_gstride); print("\n");
print("tma_gstride : "); print(tma_gstride); print("\n");
print("gbasis : "); print(gbasis); print("\n");
print("tile_gbasis : "); print(tile_gbasis); print("\n");
print("tma_gbasis : "); print(tma_gbasis); print("\n");
#endif
//
// TMA desc creation
//
constexpr int tma_dim = decltype(rank(tma_gbasis))::value;
//
// TMA gmem desc info
//
void* gmem_address = (void*) raw_pointer_cast(gtensor_T.data());
auto gmem_layout = gtensor_T.layout();
cute::array<uint64_t, 5> gmem_prob_shape = {1,1,1,1,1};
cute::array<uint64_t, 5> gmem_prob_stride = {0,0,0,0,0};
// Use the indirections in tma_gbasis in the values of flat_glayout to construct the gmem shapes/strides
for_each(make_seq<tma_dim>{}, [&](auto i) {
for_each(stride<i>(tma_gbasis), [&](auto ej) {
// Problem stride
uint64_t stride_j = ceil_div(basis_get(ej, stride(gmem_layout)) * sizeof_bits_v<TmaInternalType>, 8);
uint64_t old_stride = gmem_prob_stride[i];
gmem_prob_stride[i] = gcd(gmem_prob_stride[i], stride_j);
// Problem shape
uint64_t shape_j = basis_get(ej, shape(gmem_layout));
if (gmem_prob_stride[i] != 0) {
// Recurrence: g_shape = (s_i - 1) * (d_i / gcd_j d_j) + 1
gmem_prob_shape[i] = (gmem_prob_shape[i]-1) * (old_stride / gmem_prob_stride[i])
+ (shape_j-1) * (stride_j / gmem_prob_stride[i])
+ 1;
} else {
gmem_prob_shape[i] = shape_j;
}
});
});
assert((reinterpret_cast<uint64_t>(gmem_address) & 0b1111) == 0); // Address must be 16B-aligned
assert(gmem_prob_shape[0] >= (uint64_t(1))); // Size must be min 1
assert(gmem_prob_shape[0] <= (uint64_t(1) << 32)); // Size must be max 2^32
assert(gmem_prob_shape[1] >= (uint64_t(1))); // Size must be min 1
assert(gmem_prob_shape[1] <= (uint64_t(1) << 32)); // Size must be max 2^32
assert(gmem_prob_shape[2] >= (uint64_t(1))); // Size must be min 1
assert(gmem_prob_shape[2] <= (uint64_t(1) << 32)); // Size must be max 2^32
assert(gmem_prob_shape[3] >= (uint64_t(1))); // Size must be min 1
assert(gmem_prob_shape[3] <= (uint64_t(1) << 32)); // Size must be max 2^32
assert(gmem_prob_shape[4] >= (uint64_t(1))); // Size must be min 1
assert(gmem_prob_shape[4] <= (uint64_t(1) << 32)); // Size must be max 2^32
// TMA descriptor does not store the zeroth stride and assumes it is 1 (TmaInternalType element).
assert(gmem_prob_stride[0] == sizeof(TmaInternalType) && "Majorness of smem doesn't match majorness of gmem");
assert((gmem_prob_stride[1]) < (uint64_t(1) << 40)); // Stride must be max 2^40
assert((gmem_prob_stride[1] & 0b1111) == 0); // Stride must be multiple of 16B (128b)
assert((gmem_prob_stride[2]) < (uint64_t(1) << 40)); // Stride must be max 2^40
assert((gmem_prob_stride[2] & 0b1111) == 0); // Stride must be multiple of 16B (128b)
assert((gmem_prob_stride[3]) < (uint64_t(1) << 40)); // Stride must be max 2^40
assert((gmem_prob_stride[3] & 0b1111) == 0); // Stride must be multiple of 16B (128b)
assert((gmem_prob_stride[4]) < (uint64_t(1) << 40)); // Stride must be max 2^40
assert((gmem_prob_stride[4] & 0b1111) == 0); // Stride must be multiple of 16B (128b)
//
// TMA smem desc info
//
cute::array<uint32_t, 5> smem_box_shape = {1,1,1,1,1};
cute::array<uint32_t, 5> smem_box_stride = {1,1,1,1,1};
// The smem box is simply given by the sizes of the modes in tma_gbasis
for_each(make_seq<tma_dim>{}, [&](auto i) {
smem_box_shape[i] *= size<i>(tma_gbasis);
});
assert(smem_box_shape[0] >= (uint32_t(1))); // Size must be min 1
assert(smem_box_shape[0] <= (uint32_t(1) << 8)); // Size must be max 2^8 = 256
assert(smem_box_shape[1] >= (uint32_t(1))); // Size must be min 1
assert(smem_box_shape[1] <= (uint32_t(1) << 8)); // Size must be max 2^8 = 256
assert(smem_box_shape[2] >= (uint32_t(1))); // Size must be min 1
assert(smem_box_shape[2] <= (uint32_t(1) << 8)); // Size must be max 2^8 = 256
assert(smem_box_shape[3] >= (uint32_t(1))); // Size must be min 1
assert(smem_box_shape[3] <= (uint32_t(1) << 8)); // Size must be max 2^8 = 256
assert(smem_box_shape[4] >= (uint32_t(1))); // Size must be min 1
assert(smem_box_shape[4] <= (uint32_t(1) << 8)); // Size must be max 2^8 = 256
assert(smem_box_stride[0] >= (uint32_t(1))); // Stride must be min 1
assert(smem_box_stride[0] <= (uint32_t(8))); // Stride must be max 2^3 = 8
assert(smem_box_stride[1] >= (uint32_t(1))); // Stride must be min 1
assert(smem_box_stride[1] <= (uint32_t(8))); // Stride must be max 2^3 = 8
assert(smem_box_stride[2] >= (uint32_t(1))); // Stride must be min 1
assert(smem_box_stride[2] <= (uint32_t(8))); // Stride must be max 2^3 = 8
assert(smem_box_stride[3] >= (uint32_t(1))); // Stride must be min 1
assert(smem_box_stride[3] <= (uint32_t(8))); // Stride must be max 2^3 = 8
assert(smem_box_stride[4] >= (uint32_t(1))); // Stride must be min 1
assert(smem_box_stride[4] <= (uint32_t(8))); // Stride must be max 2^3 = 8
//
// Construct the descriptor
//
TmaDescriptor tma_desc = {0};
//
// TMA general info
//
#if (__CUDACC_VER_MAJOR__ >= 12) && !defined(__CUDACC_RTC__)
CUtensorMapDataType tma_format = TMA::to_CUtensorMapDataType<TmaInternalType>();
CUtensorMapInterleave tma_interleave = CU_TENSOR_MAP_INTERLEAVE_NONE;
CUtensorMapL2promotion tma_l2Promotion = CU_TENSOR_MAP_L2_PROMOTION_L2_128B;
CUtensorMapFloatOOBfill tma_oobFill = CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE;
// TMA smem swizzle type
CUtensorMapSwizzle smem_swizzle = TMA::to_CUtensorMapSwizzle(get_tma_swizzle_bits(swizzle));
CUresult result = cuTensorMapEncodeTiled(
&tma_desc,
tma_format,
tma_dim,
gmem_address,
gmem_prob_shape.data(),
gmem_prob_stride.data() + 1, // gmem_prob_stride[0] implicitly 1
smem_box_shape.data(),
smem_box_stride.data(),
tma_interleave,
smem_swizzle,
tma_l2Promotion,
tma_oobFill);
if (result != CUDA_SUCCESS) {
std::cerr << "TMA Desc Addr: " << &tma_desc
<< "\nformat " << tma_format
<< "\ndim " << tma_dim
<< "\ngmem_address " << gmem_address
<< "\nglobalDim " << gmem_prob_shape
<< "\nglobalStrides " << gmem_prob_stride
<< "\nboxDim " << smem_box_shape
<< "\nelementStrides " << smem_box_stride
<< "\ninterleave " << tma_interleave
<< "\nswizzle " << smem_swizzle
<< "\nl2Promotion " << tma_l2Promotion
<< "\noobFill " << tma_oobFill << std::endl;
std::cerr << "Error: Failed to initialize the TMA descriptor " << result << std::endl;
assert(false);
}
#endif // (__CUDACC_VER_MAJOR__ >= 12) && !defined(__CUDACC_RTC__)
auto recast_ratio = cute::ratio(Int<sizeof_bits<typename GEngine::value_type>::value>{},
Int<sizeof_bits< TmaInternalType>::value>{});
// Finally, get the inverse permutation of the E<i> bases for the mocked gmem stride
// NOTE This is essentially ArithmeticTuple inverse...
auto gmem_stride_bases = transform_leaf(stride(gbasis), [&](auto ei) {
auto si = basis_get(ei, shape(gmem_layout));
auto di = basis_get(ei, stride(gmem_layout));
if constexpr (is_constant<1, decltype(si)>::value || is_constant<0, decltype(di)>::value) {
return Int<0>{}; // If size-1 or stride-0, return arithmetic identity -- no contribution to the TMA
} else {
auto tma_gbasis_stride = stride(tma_gbasis);
// Find j such that E<i> is in stride<j>(tma_gbasis)
using EI = decltype(ei);
[[maybe_unused]] auto j = find_if(tma_gbasis_stride, [&](auto tma_stride_j) { return any_of(tma_stride_j, [&](auto dj) { return dj == EI{}; }); });
if constexpr (decltype(j == rank(tma_gbasis_stride))::value) {
return Int<0>{}; // If not-found, return arithmetic identity -- no contribution to the TMA
} else
if constexpr (decltype(j == Int<0>{})::value) {
auto scale = recast_ratio * basis_get(ei, stride(gtensor));
return E<j>{} * scale; // Return TMA Coord basis -- with a recast scale factor
} else
if constexpr (decltype(rank<j>(tma_gbasis_stride) == Int<1>{})::value) {
return E<j>{}; // Return TMA Coord basis -- known scale of Int<1>{}
} else {
int32_t scale = ceil_div(int32_t(di * sizeof_bits_v<TmaInternalType> / cute::max(gmem_prob_stride[j], 16)), 8);
return E<j>{} * scale; // Return TMA Coord basis -- with a dynamic scale factor
}
}
});
#if 0
print("tma_gbasis : "); print(gmem_stride_bases); print("\n");
#endif
using AuxParams = AuxTmaParams<decltype(gmem_stride_bases),
decltype(tma_gbasis),
decltype(swizzle)>;
return cute::make_tuple(tma_desc, AuxParams{gmem_stride_bases});
}
// The "logical TMA tid" is a map from the CTA rank to its logical id
// within the instruction. It works like a mask or ordering on the
// CTAs. For non-multicast TMA, all CTAs should map to 0. For
// multicast TMA of size 4, CTAs will be mapped to {0,1,2,3}.
template <class TmaInternalType,
class CopyOp,
class GEngine, class GLayout,
class SLayout,
class TShape, class TStride,
class VShape, class VStride>
CUTE_HOST_RTC
auto
make_tma_copy_tiled(CopyOp,
Tensor<GEngine,GLayout> const& gtensor, // Full GMEM Tensor
SLayout const& slayout, // CTA Tile of SMEM
Layout<TShape,TStride> const& cta_t_map, // T: CTA thr idx -> logical TMA tid
Layout<VShape,VStride> const& cta_v_map) // V: CTA val idx -> gmem mode
{
//
// TMA parameter checking
//
CUTE_STATIC_ASSERT_V(product_each(shape(slayout)) == product_each(shape(cta_v_map)),
"TMA requires CTA_Tile and SLayout top-level shape equivalence.");
CUTE_STATIC_ASSERT_V(size(slayout) % cosize(cta_t_map) == Int<0>{},
"Number of active CTAs in TMA must divide domain size of slayout.");
//
// TMA slayout manipulation
//
// Invert the smem to get the largest contiguous vector in the smem layout
// smem idx -> smem coord
auto inv_smem_layout = right_inverse(get_nonswizzle_portion(slayout));
// Compose with the V-Map to convert smem coord (CTA val idx) to gmem mode
// smem idx -> gmem mode
auto sidx_to_gmode = coalesce(composition(cta_v_map, inv_smem_layout));
#if 0
print("g_tensor : "); print(gtensor); print("\n");
print("s_layout : "); print(slayout); print("\n");
print("cta_t_map : "); print(cta_t_map); print("\n");
print("cta_v_map : "); print(cta_v_map); print("\n");
print("inv_s_layout : "); print(inv_smem_layout); print("\n");
print("sidx_to_gmode : "); print(sidx_to_gmode); print("\n");
#endif
//
// TMA gtensor manipulation
//
// Generate a TupleBasis for the gtensor
// gmem coord -> gmem coord
auto glayout_basis = make_identity_layout(shape(gtensor));
// Tile the modes of gtensor with the truncated cta_v_map o inv_smem_layout_trunc
// smem idx -> gmem coord
auto tma_layout_full = flatten(composition(glayout_basis, sidx_to_gmode));
// Truncate any incompatibilities -- no starting in the middle of gmodes
auto smem_rank = find_if(stride(tma_layout_full), [](auto e) {
[[maybe_unused]] auto v = basis_value(e);
return not is_constant<1,decltype(v)>{};
});
static_assert(smem_rank > 0, "Could not find a common tile-gmem vectorization. Does the Tile select out major GMEM modes?");
// Keep only the static-1 basis modes into gmem
auto tma_layout_trunc = take<0,smem_rank>(tma_layout_full);
// Keep only the portion each multicast CTA will be responsible for
auto tma_layout_v = composition(tma_layout_trunc, shape_div(size(tma_layout_trunc), cosize(cta_t_map)));
#if 0
print("glayout_basis : "); print(glayout_basis); print("\n");
print("tma_layout_full : "); print(tma_layout_full); print("\n");
print("tma_layout_trunc: "); print(tma_layout_trunc); print("\n");
print("tma_layout_v : "); print(tma_layout_v); print("\n");
#endif
//
// Construct the TMA Desc and the strides of the TMA Tensor
//
auto [tma_desc, aux_params] = detail::make_tma_copy_desc<TmaInternalType>(gtensor,
tma_layout_v,
get_swizzle_portion(slayout));
//
// Construct the Copy_Traits
//
using T = typename GEngine::value_type;
constexpr int num_bits_per_tma = decltype(size(tma_layout_trunc))::value * sizeof_bits_v<T>;
using Traits = Copy_Traits<CopyOp, cute::C<num_bits_per_tma>, decltype(aux_params)>;
using Atom = Copy_Atom<Traits, T>;
Traits tma_traits{tma_desc, aux_params};
#if 0
print("num_bits_per_tma : "); print(num_bits_per_tma); print("\n");
print("g_stride_bases : "); print(tma_traits.aux_params_.g_stride_); print("\n");
#endif
//
// Construct the TiledCopy
//
auto cta_tiler = product_each(shape(cta_v_map));
// CTA V -> smem_coord
auto layout_v = composition(inv_smem_layout, size(tma_layout_trunc));
auto layout_V = tile_to_shape(make_layout(layout_v), size(cta_v_map));
// CTA T -> smem idx
auto layout_t = make_layout(cosize(cta_t_map), shape_div(size(tma_layout_trunc), cosize(cta_t_map)));
// CTA TID -> smem coord
auto layout_T = composition(inv_smem_layout, composition(layout_t, cta_t_map));
// Combine with the T mapping
auto layout_TV = make_layout(layout_T, layout_V);
#if 0
print("cta_tiler : "); print(cta_tiler); print("\n");
print("layout_VT : "); print(layout_VT); print("\n");
print("layout_TV : "); print(layout_TV); print("\n");
#endif
return TiledCopy<Atom, decltype(layout_TV), decltype(cta_tiler)>{tma_traits};
}
} // end namespace detail
/** Make a CuTe CTA-collective TiledCopy for a TMA operation.
*
* @param CopyOp The target copy operation: SM90_TMA_LOAD, SM90_TMA_LOAD_MULTICAST, SM90_TMA_STORE
* @param gtensor The GMEM Tensor to be involved in the TMA.
* @param slayout The SMEM Layout to be involved in the TMA.
* @param cta_tile The CTA-local tile that each CTA will be tiling GMEM with.
* This is often the blk_shape that is used to tile the GMEM for CTAs:
* local_tile(gtensor, blk_shape, blk_coord) -> CTA-local tile of gtensor
* @param cluster_size When using SM90_TMA_LOAD_MULTICAST, this can be a (static) power-of-2 <= 16
* defining the multicast size (used to further partition the SMEM)
* Else, static-1
*
* This code attempts to maximize the TMA box size. It does this by tracing
* the SMEM "vector" -- the inverse of the smem layout -- to find the largest
* contiguous array of smem that can be written to/from global memory given
* the constraints that the TMA instruction imposes.
*
* This is accomplished by assigning "basis" strides to the GMEM to track which
* modes of SMEM map to which modes of GMEM, then reorder the modes of GMEM according
* to the SMEM vector, and then using those GMEM/SMEM modes to fill in the desc.
*
* Examples:
using T = float;
T* gptr = nullptr;
{
// Simple 2D
Tensor gtensor = make_tensor(gptr, make_shape(1024, 256), GenRowMajor{}); // K-Major GMEM
auto slayout = make_layout(make_shape(_64{}, _32{}), GenRowMajor{}); // K-Major SMEM
auto tma = make_tma_copy(SM90_TMA_LOAD{}, gtensor, slayout);
}
{
// GMMA 2D
Tensor gtensor = make_tensor(gptr, make_shape(1024, 256)); // MN-Major GMEM
auto slayout = tile_to_shape(GMMA::Layout_MN_SW128_Atom<T>{}, make_shape(_128{},_64{})); // MN-Major Swizzled+Tiled 128x64 SMEM
auto tma = make_tma_copy(SM90_TMA_LOAD{}, gtensor, slayout);
}
{
// 3D
Tensor gtensor = make_tensor(gptr, make_shape(1024, 32, 512), make_stride(64, Int<1>{}, 65536)); // GMEM
auto slayout = make_layout(make_shape(_16{}, _8{}, _2{}), make_stride(_16{}, _1{}, _8{})); // SMEM w/ same major-mode
auto tma = make_tma_copy(SM90_TMA_LOAD{}, gtensor, slayout);
}
{
// cuTENSOR 4D
auto layout = make_shape(make_shape(32,40),make_shape(make_shape(8,8),656)); // GMEM
auto cta_tile = make_shape(_128{},make_shape(_32{},_2{})); // GMEM Tiling:
// Take 128-elem from m: m0 must divide 128,
// m-last may be predicated
// Take 32-elem from k0, 2-elem from k1
auto slayout = make_layout(cta_tile); // Col-Major SMEM
auto tma = make_tma_copy(SM90_TMA_LOAD{}, gtensor, slayout, cta_tile, Int<1>{});
}
*
* Check the TMA box size and desc:
print("TMA Box size: "); print(typename decltype(tma)::Tiler_MN{}); print("\n");
print("TMA desc : "); print(tma.tma_desc_); print("\n");
*
* Usage:
Tensor mA = tma_a.get_tma_tensor(make_shape(M,N)); // (M,N) TMA coord tensor
Tensor gA = local_tile(mA, cta_tile, cta_coord); // (BLK_M,BLK_N) TMA coord tensor for this CTA
Tensor sA = make_tensor(make_smem_ptr<T>(sptr), slayout); // (BLK_M,BLK_N) SMEM tensor
auto cta_tma = tma.get_slice(cta_idx_in_cluster); // Slice for multicast partitioning
Tensor tAgA = cta_tma.partition_S(gA); // Partition for src
Tensor tAsA = cta_tma.partition_D(sA); // Partition for dst
copy(tma.with(barrier, mcast_mask), tAgA, tAsA); // copy with supporting TMA params
*/
template <class TmaInternalType,
class CopyOp,
class GEngine, class GLayout,
class SLayout,
class CTA_Tiler,
class Cluster_Size>
CUTE_HOST_RTC
auto
make_tma_copy(CopyOp const& copy_op,
Tensor<GEngine,GLayout> const& gtensor,
SLayout const& slayout,
CTA_Tiler const& cta_tiler,
Cluster_Size const& cluster_size)
{
auto cta_v_tile = make_identity_layout(shape(gtensor)).compose(cta_tiler);
auto cta_t_tile = make_layout(cluster_size);
return detail::make_tma_copy_tiled<TmaInternalType>(copy_op,
gtensor,
slayout,
cta_t_tile,
cta_v_tile);
}
// Explicit defaulting
template <class CopyOp,
class GEngine, class GLayout,
class SLayout,
class CTA_Tile,
class Cluster_Size>
CUTE_HOST_RTC
auto
make_tma_copy(CopyOp const& copy_op,
Tensor<GEngine,GLayout> const& gtensor,
SLayout const& slayout,
CTA_Tile const& cta_tile,
Cluster_Size const& cluster_size)
{
using TmaInternalType = typename GEngine::value_type;
return make_tma_copy<TmaInternalType>(copy_op,
gtensor,
slayout,
cta_tile,
cluster_size);
}
template <class CopyOp,
class GEngine, class GLayout,
class SLayout>
CUTE_HOST_RTC
auto
make_tma_copy(CopyOp const& copy_op,
Tensor<GEngine,GLayout> const& gtensor,
SLayout const& slayout)
{
return make_tma_copy(copy_op, gtensor, slayout, product_each(shape(slayout)), Int<1>{});
}
template <class CopyOp,
class GEngine, class GLayout,
class SLayout,
class Cluster_Size>
CUTE_HOST_RTC
auto
make_tma_copy(CopyOp const& copy_op,
Tensor<GEngine,GLayout> const& gtensor,
SLayout const& slayout,
Cluster_Size const& cluster_size)
{
return make_tma_copy(copy_op, gtensor, slayout, product_each(shape(slayout)), cluster_size);
}
} // end namespace cute
Reference in New Issue View Git Blame Copy Permalink