ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_BUILDER] Convolution forward transfer concepts. (#3535)

Browse Source 
* Rename member variable to better reflect its actuall meaning.

* Add transfer checks for conv fwd xdl.

* Validate tensor layouts & vector size conv fwd v3.

* Add combined transfer concepts.

* Add transfer concepts for conv fwd factories.

* Fix clang format

* Add helper instruction to get max mem vector instruction width.

* Apply review comments.

* Rename thread cluster access(->arrange) order concept

* FIx merge artifacts.

* Add generic access order limits into block transfer concept.
This commit is contained in:
Adam Osewski
2026-01-19 10:54:10 +01:00
committed by GitHub
parent fe40a5d139
commit 1a6d1b59ef
13 changed files with 570 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files

192
experimental/builder/test/utils/ckb_conv_test_configs.hpp
View File

@@ -53,25 +53,25 @@ constexpr DlTransfer<5> DlTransfer5D{.a = DlBlockTransfer_1x8x1x1x1,
constexpr Transfer<> Transfer_4x64x1{
.a =
{
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = false},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 4,
.is_direct_load = false,
.lds_padding = false},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.b =
{
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = false},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 4,
.lds_dst_scalar_per_vector = 4,
.is_direct_load = false,
.lds_padding = false},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.c =
{
@@ -86,25 +86,25 @@ constexpr Transfer<> Transfer_4x64x1{
constexpr Transfer<4> BwdTransfer_4x64x1{
.a =
{
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1, .k_batch_size = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 4,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {0, 3, 1, 2},
.src_access_order = {0, 2, 1, 3},
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1, .k_batch_size = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 4,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {0, 3, 1, 2},
.src_access_order = {0, 2, 1, 3},
},
.b =
{
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1, .k_batch_size = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 4,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {0, 3, 1, 2},
.src_access_order = {0, 2, 1, 3},
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1, .k_batch_size = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 4,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {0, 3, 1, 2},
.src_access_order = {0, 2, 1, 3},
},
.c =
{
@@ -119,25 +119,25 @@ constexpr Transfer<4> BwdTransfer_4x64x1{
constexpr Transfer<> BwdTransfer_4x8x1_4x16x1_v3{
.a =
{
.block_transfer = {.k0 = 4, .m_n = 8, .k1 = 1},
.lds_transfer = {.src_vector_dim = 1,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 2,
.is_direct_load = false,
.lds_padding = false},
.block_transfer_access_order = {2, 0, 1},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 8, .k1 = 1},
.lds_transfer = {.src_vector_dim = 1,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 2,
.is_direct_load = false,
.lds_padding = false},
.thread_cluster_arrange_order = {2, 0, 1},
.src_access_order = {1, 0, 2},
},
.b =
{
.block_transfer = {.k0 = 4, .m_n = 16, .k1 = 1},
.lds_transfer = {.src_vector_dim = 1,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 2,
.is_direct_load = false,
.lds_padding = false},
.block_transfer_access_order = {2, 0, 1},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 16, .k1 = 1},
.lds_transfer = {.src_vector_dim = 1,
.src_scalar_per_vector = 2,
.lds_dst_scalar_per_vector = 2,
.is_direct_load = false,
.lds_padding = false},
.thread_cluster_arrange_order = {2, 0, 1},
.src_access_order = {1, 0, 2},
},
.c =
{
@@ -152,25 +152,25 @@ constexpr Transfer<> BwdTransfer_4x8x1_4x16x1_v3{
constexpr Transfer<> Transfer_4x64x1_fp8{
.a =
{
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.b =
{
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 64, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.c =
{
@@ -185,25 +185,25 @@ constexpr Transfer<> Transfer_4x64x1_fp8{
constexpr Transfer<> Transfer_4x16x1{
.a =
{
.block_transfer = {.k0 = 4, .m_n = 16, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 16, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.b =
{
.block_transfer = {.k0 = 4, .m_n = 16, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 16, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 8,
.lds_dst_scalar_per_vector = 8,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.c =
{
@@ -219,25 +219,25 @@ constexpr Transfer<> Transfer_4x16x1{
constexpr Transfer<> Transfer_4x32x1{
.a =
{
.block_transfer = {.k0 = 4, .m_n = 32, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 16,
.lds_dst_scalar_per_vector = 16,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 32, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 16,
.lds_dst_scalar_per_vector = 16,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.b =
{
.block_transfer = {.k0 = 4, .m_n = 32, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 16,
.lds_dst_scalar_per_vector = 16,
.is_direct_load = false,
.lds_padding = true},
.block_transfer_access_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
.block_transfer = {.k0 = 4, .m_n = 32, .k1 = 1},
.lds_transfer = {.src_vector_dim = 2,
.src_scalar_per_vector = 16,
.lds_dst_scalar_per_vector = 16,
.is_direct_load = false,
.lds_padding = true},
.thread_cluster_arrange_order = {1, 0, 2},
.src_access_order = {1, 0, 2},
},
.c =
{

2
experimental/builder/test/utils/conv_algorithm_type_utils.hpp
View File

@@ -165,7 +165,7 @@ template <size_t N = 3>
inline std::string to_string(InputTransfer<N> t)
{
std::ostringstream oss;
oss << to_string(t.block_transfer) << "," << to_string(t.block_transfer_access_order) << ","
oss << to_string(t.block_transfer) << "," << to_string(t.thread_cluster_arrange_order) << ","
<< to_string(t.src_access_order) << "," << t.lds_transfer.src_vector_dim << ","
<< t.lds_transfer.src_scalar_per_vector << "," << t.lds_transfer.lds_dst_scalar_per_vector
<< "," << (t.lds_transfer.lds_padding ? "true" : "false");