ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-01 20:21:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "[CK-Tile] Add the API to load SGPR (#2878)" (#2904)

Browse Source 
This reverts commit 2cbbf5dcb3.
This commit is contained in:
asleepzzz
2025-09-24 05:33:51 +08:00
committed by GitHub
parent 959df2a155
commit f161b5b738
40 changed files with 167 additions and 273 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
include/ck_tile/ops/add_rmsnorm2d_rdquant/pipeline/add_rmsnorm2d_rdquant_fwd_pipeline_three_pass.hpp
View File

@@ -92,13 +92,13 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
index_t num_n_tile_iteration =
amd_wave_read_first_lane(integer_divide_ceil(row_size, Block_N));
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
using XTensorType = decltype(cast_tile<ComputeDataType>(load_tile(a_window)));
auto square_sum = block_reduce2d.template MakeYBlockTile<XTensorType>();
set_tile(square_sum, reduce_square_sum_func.GetIdentityValue<ComputeDataType>());
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto a = load_tile(a_window);
const auto b = load_tile(b_window);
@@ -149,7 +149,7 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
if constexpr(kSaveX)
__syncthreads();
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = [&]() {
if constexpr(kSaveX)
@@ -226,7 +226,7 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
}
move_tile_window(gamma_window, {Block_N});
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = [&]() {
if constexpr(kSaveX)

6
include/ck_tile/ops/batched_transpose/kernel/batched_transpose_kernel.hpp
View File

@@ -84,9 +84,9 @@ struct BatchedTransposeKernel
static constexpr ck_tile::index_t VectorSizeOutput = Problem::VectorSizeOutput;
static constexpr ck_tile::index_t VectorStrideOutput = 1;
const auto iM = amd_wave_read_first_lane(blockIdx.x * kMPerBlock);
const auto iN = amd_wave_read_first_lane(blockIdx.y * kNPerBlock);
const auto offset = amd_wave_read_first_lane(blockIdx.z * kargs.height * kargs.width);
const auto iM = __builtin_amdgcn_readfirstlane(blockIdx.x * kMPerBlock);
const auto iN = __builtin_amdgcn_readfirstlane(blockIdx.y * kNPerBlock);
const auto offset = __builtin_amdgcn_readfirstlane(blockIdx.z * kargs.height * kargs.width);
const auto x_m_n = [&]() {
const auto x_dram_naive = make_naive_tensor_view<address_space_enum::global>(

4
include/ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp
View File

@@ -598,8 +598,8 @@ struct FlatmmKernel
CK_TILE_DEVICE void operator()(KernelArgs kargs) const
{
const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockIdx.x);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const SplitKBatchOffset splitk_batch_offset(kargs);
// options

4
include/ck_tile/ops/fmha/kernel/fmha_batch_prefill_kernel.hpp
View File

@@ -707,8 +707,8 @@ struct FmhaBatchPrefillWithPagedKVCacheKernel
// divide problem
const auto [i_tile_m, i_tile_n, i_nhead, i_batch] = GetTileIndex(kargs);
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = amd_wave_read_first_lane(i_tile_n * FmhaPipeline::kN1);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
long_index_t batch_offset_q = 0;
long_index_t batch_offset_bias = 0;

6
include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp
View File

@@ -690,7 +690,7 @@ struct FmhaBwdDQDKDVKernel
// divide problem
const auto [i_tile_n, i_nhead, i_batch] = GetTileIndex();
const index_t i_n0 = amd_wave_read_first_lane(i_tile_n * FmhaPipeline::kN0);
const index_t i_n0 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN0);
long_index_t batch_offset_q = 0;
long_index_t batch_offset_k = 0;
@@ -1338,7 +1338,7 @@ struct FmhaBwdOGradDotOKernel
// divide problem
const auto [i_tile_m, i_nhead, i_batch] = GetTileIndex();
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * kM0);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * kM0);
long_index_t batch_offset_o = 0;
long_index_t batch_offset_do = 0;
@@ -1618,7 +1618,7 @@ struct FmhaBwdConvertQGradKernel
// divide problem
const auto [i_tile_m, i_nhead, i_batch] = GetTileIndex();
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * kM0);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * kM0);
long_index_t batch_offset_dq = 0;
long_index_t batch_offset_dq_acc = 0;

4
include/ck_tile/ops/fmha/kernel/fmha_fwd_appendkv_kernel.hpp
View File

@@ -262,8 +262,8 @@ struct FmhaFwdAppendKVKernel
// divide problem
const auto [i_tile, i_nhead, i_batch] = GetTileIndex(kargs);
const index_t i_m0 = amd_wave_read_first_lane(i_tile * FmhaPipeline::kM0);
const index_t i_n0 = amd_wave_read_first_lane(i_tile * FmhaPipeline::kN0);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile * FmhaPipeline::kM0);
const index_t i_n0 = __builtin_amdgcn_readfirstlane(i_tile * FmhaPipeline::kN0);
const index_t i_cache_batch = [&, i_batch_ = i_batch] {
if constexpr(kIsPagedKV)

4
include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp
View File

@@ -1060,8 +1060,8 @@ struct FmhaFwdKernel
// divide problem
const auto [i_tile_m, i_tile_n, i_nhead, i_batch] = GetTileIndex(kargs);
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = amd_wave_read_first_lane(i_tile_n * FmhaPipeline::kN1);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
long_index_t batch_offset_q = 0;
long_index_t batch_offset_k = 0;

4
include/ck_tile/ops/fmha/kernel/fmha_fwd_pagedkv_kernel.hpp
View File

@@ -880,8 +880,8 @@ struct FmhaFwdPagedKVKernel
// divide problem
const auto [i_tile_m, i_tile_n, i_nhead, i_batch] = GetTileIndex(kargs);
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = amd_wave_read_first_lane(i_tile_n * FmhaPipeline::kN1);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
long_index_t batch_offset_q = 0;
long_index_t batch_offset_k = 0;

4
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp
View File

@@ -281,8 +281,8 @@ struct FmhaFwdSplitKVCombineKernel
// divide problem
const auto [i_tile_m, i_tile_n, i_nhead, i_batch] = GetTileIndex(kargs);
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = amd_wave_read_first_lane(i_tile_n * FmhaPipeline::kN1);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
long_index_t batch_offset_lse_acc = 0;
long_index_t batch_offset_o_acc = 0;

4
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp
View File

@@ -589,8 +589,8 @@ struct FmhaFwdSplitKVKernel
// divide problem
const auto [i_tile_m, i_tile_n, i_split, i_nhead, i_batch] = GetTileIndex(kargs);
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = amd_wave_read_first_lane(i_tile_n * FmhaPipeline::kN1);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
long_index_t batch_offset_q = 0;
long_index_t batch_offset_k = 0; // unused for paged-kvcache

4
include/ck_tile/ops/fmha/kernel/fmha_fwd_v3_kernel.hpp
View File

@@ -361,8 +361,8 @@ struct FmhaFwdV3Kernel
// divide problem
const auto [i_tile_m, i_tile_n, i_nhead, i_batch] = GetTileIndex(kargs);
const index_t i_m0 = amd_wave_read_first_lane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = amd_wave_read_first_lane(i_tile_n * FmhaPipeline::kN1);
const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
long_index_t batch_offset_q = 0;
long_index_t batch_offset_k = 0;

4
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_pagedkv_pipeline_qr_ks_vs.hpp
View File

@@ -320,9 +320,9 @@ struct BlockFmhaFwdPagedKVPipelineQRKSVS
k_block_tile = load_tile(k_dram_window);
}
auto physical_next_block_id_k =
amd_wave_read_first_lane(k_page_block_navigator.prefetch_table_id(
__builtin_amdgcn_readfirstlane(k_page_block_navigator.prefetch_table_id(
i_page_block_k, k_dram_block_window, {kN0, 0}));
auto physical_next_block_id_v = amd_wave_read_first_lane(
auto physical_next_block_id_v = __builtin_amdgcn_readfirstlane(
v_page_block_navigator.prefetch_table_id(i_page_block_v, v_dram_window, {0, kK1}));
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)

6
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs.hpp
View File

@@ -321,9 +321,9 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
k_block_tile = load_tile(k_dram_window);
}
auto physical_next_block_id_k =
amd_wave_read_first_lane(k_page_block_navigator.prefetch_table_id(
__builtin_amdgcn_readfirstlane(k_page_block_navigator.prefetch_table_id(
i_page_block_k, k_dram_block_window, {kN0, 0}));
auto physical_next_block_id_v = amd_wave_read_first_lane(
auto physical_next_block_id_v = __builtin_amdgcn_readfirstlane(
v_page_block_navigator.prefetch_table_id(i_page_block_v, v_dram_window, {0, kK1}));
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
@@ -618,7 +618,7 @@ struct BlockFmhaFwdSplitKVPipelineQRKSVS
&i_page_block_v_ = i_page_block_v,
&v_dram_window_ = v_dram_window](auto i_k1) {
auto physical_next_block_id_v_ =
amd_wave_read_first_lane(v_page_block_navigator.prefetch_table_id(
__builtin_amdgcn_readfirstlane(v_page_block_navigator.prefetch_table_id(
i_page_block_v_, v_dram_window_, {0, kK1}));
const auto v = load_tile(v_dram_window_); // load next v
block_sync_lds();

8
include/ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp
View File

@@ -240,7 +240,7 @@ struct FusedMoeGemmKernel
if constexpr(UseUK)
{
__shared__ CK_TILE_LDS_ADDR char smem[GetSmemSize()];
IndexDataType num_sorted_tiles = amd_wave_read_first_lane(
IndexDataType num_sorted_tiles = __builtin_amdgcn_readfirstlane(
*reinterpret_cast<const IndexDataType*>(kargs.num_sorted_tiles_ptr));
num_sorted_tiles = num_sorted_tiles / BlockShape::Block_M0;
@@ -261,7 +261,7 @@ struct FusedMoeGemmKernel
{
// allocate LDS
// __shared__ char smem_ptr[GetSmemSize()];
IndexDataType num_sorted_tiles = amd_wave_read_first_lane(
IndexDataType num_sorted_tiles = __builtin_amdgcn_readfirstlane(
*reinterpret_cast<const IndexDataType*>(kargs.num_sorted_tiles_ptr));
constexpr index_t hidden_radio_0 = IsGateOnly ? 1 : 2;
@@ -283,14 +283,14 @@ struct FusedMoeGemmKernel
return;
const IndexDataType expert_id =
amd_wave_read_first_lane(reinterpret_cast<const IndexDataType*>(
__builtin_amdgcn_readfirstlane(reinterpret_cast<const IndexDataType*>(
kargs.sorted_expert_ids_ptr)[sorted_tile_id]);
// index along intermediate_size
// index_t hidden_idx = __builtin_amdgcn_readfirstlane(intermediate_tile_id *
// BlockShape::Block_N0);
index_t interm_idx_nr =
amd_wave_read_first_lane(intermediate_tile_id * BlockShape::Block_Nr0);
__builtin_amdgcn_readfirstlane(intermediate_tile_id * BlockShape::Block_Nr0);
const auto a_coord = Pipeline::GetACoord(); // 2d thread offset, [i_row, i_col]
const auto sorted_token_id =

2
include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp
View File

@@ -756,7 +756,7 @@ struct MoeSortingKernel
void* smem) const
{
const index_t tid = static_cast<index_t>(threadIdx.x);
const index_t wid = amd_wave_read_first_lane(tid / get_warp_size());
const index_t wid = __builtin_amdgcn_readfirstlane(tid / get_warp_size());
const index_t lid = __lane_id();
constexpr index_t block_size = 256; // blockDim.x;
const index_t sub_tokens = smem_rows - 2; // sub_tokens_mdiv.divisor;

6
include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp
View File

@@ -184,17 +184,17 @@ struct FusedMoeGemmPipeline_FlatmmUk
index_t nr_1 = kargs.hidden_size / BlockShape::Warp_N1;
index_t kr_1 = shared_intermediate_size_1 / BlockShape::Warp_K1;
const IndexDataType expert_id = amd_wave_read_first_lane(
const IndexDataType expert_id = __builtin_amdgcn_readfirstlane(
reinterpret_cast<const IndexDataType*>(kargs.sorted_expert_ids_ptr)[sorted_tile_id]);
index_t expert_stride_0 = shared_intermediate_size_0 * kargs.hidden_size;
index_t expert_stride_1 = shared_intermediate_size_1 * kargs.hidden_size;
// nr*kr*w
index_t interm_idx_nr0 = amd_wave_read_first_lane(
index_t interm_idx_nr0 = __builtin_amdgcn_readfirstlane(
intermediate_tile_id *
BlockShape::Block_Nr0); // intermediate_tile_id * Block_N / (N in W)
index_t interm_idx_kr1 = amd_wave_read_first_lane(
index_t interm_idx_kr1 = __builtin_amdgcn_readfirstlane(
intermediate_tile_id *
BlockShape::Block_Kr1); // intermediate_tile_id * Block_N / (N in W)

20
include/ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp
View File

@@ -169,27 +169,27 @@ struct BatchedGemmKernel
CK_TILE_DEVICE void operator()(BatchedGemmKernelArgs kargs) const
{
const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockIdx.x);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const auto i_batch = amd_wave_read_first_lane(blockIdx.y);
const auto i_splitk = amd_wave_read_first_lane(blockIdx.z);
const auto i_batch = __builtin_amdgcn_readfirstlane(blockIdx.y);
const auto i_splitk = __builtin_amdgcn_readfirstlane(blockIdx.z);
const typename UniversalGemmKernel::SplitKBatchOffset splitk_batch_offset(kargs, i_splitk);
// options
const auto batch_stride_A = amd_wave_read_first_lane(kargs.batch_stride_A);
const auto batch_offset_A = amd_wave_read_first_lane(i_batch * batch_stride_A);
const auto batch_stride_A = __builtin_amdgcn_readfirstlane(kargs.batch_stride_A);
const auto batch_offset_A = __builtin_amdgcn_readfirstlane(i_batch * batch_stride_A);
const ADataType* a_ptr = static_cast<const ADataType*>(kargs.as_ptr[0]) + batch_offset_A +
splitk_batch_offset.as_k_split_offset[0];
const auto batch_stride_B = amd_wave_read_first_lane(kargs.batch_stride_B);
const auto batch_offset_B = amd_wave_read_first_lane(i_batch * batch_stride_B);
const auto batch_stride_B = __builtin_amdgcn_readfirstlane(kargs.batch_stride_B);
const auto batch_offset_B = __builtin_amdgcn_readfirstlane(i_batch * batch_stride_B);
const BDataType* b_ptr = static_cast<const BDataType*>(kargs.bs_ptr[0]) + batch_offset_B +
splitk_batch_offset.bs_k_split_offset[0];
const auto batch_stride_E = amd_wave_read_first_lane(kargs.batch_stride_E);
const auto batch_offset_C = amd_wave_read_first_lane(i_batch * batch_stride_E);
const auto batch_stride_E = __builtin_amdgcn_readfirstlane(kargs.batch_stride_E);
const auto batch_offset_C = __builtin_amdgcn_readfirstlane(i_batch * batch_stride_E);
CDataType* c_ptr = static_cast<CDataType*>(kargs.e_ptr) + batch_offset_C;
// allocate LDS

8
include/ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp
View File

@@ -73,8 +73,8 @@ struct GemmTile2DPartitioner
CK_TILE_DEVICE static auto
GetOutputTileIndex(index_t blockIdx, index_t blockIdy) noexcept -> const tuple<index_t, index_t>
{
const index_t iM = amd_wave_read_first_lane(blockIdx);
const index_t iN = amd_wave_read_first_lane(blockIdy);
const index_t iM = __builtin_amdgcn_readfirstlane(blockIdx);
const index_t iN = __builtin_amdgcn_readfirstlane(blockIdy);
return make_tuple(iM, iN);
}
};
@@ -143,8 +143,8 @@ struct GemmTile1DPartitioner
{
const index_t NBlocks = integer_divide_ceil(N_, NPerBlock);
const index_t iM = amd_wave_read_first_lane(blockIdx / NBlocks);
const index_t iN = amd_wave_read_first_lane(blockIdx - iM * NBlocks);
const index_t iM = __builtin_amdgcn_readfirstlane(blockIdx / NBlocks);
const index_t iN = __builtin_amdgcn_readfirstlane(blockIdx - iM * NBlocks);
return make_tuple(iM, iN);
}

12
include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp
View File

@@ -272,8 +272,8 @@ struct GroupedGemmKernel
const auto [iM, iN] = block_idx_2d;
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const typename Base::SplitKBatchOffset splitk_batch_offset(kargs, block_idx_z);
@@ -358,8 +358,8 @@ struct GroupedGemmKernel
const auto& d_block_window = gemm_tile_windows.at(Base::I2);
// Get hot-loop and tail configuration
const index_t num_loop =
amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const index_t num_loop = __builtin_amdgcn_readfirstlane(
TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const bool has_hot_loop = GemmPipeline::BlockHasHotloop(num_loop);
const TailNumber tail_num = GemmPipeline::GetBlockLoopTailNum(num_loop);
@@ -416,8 +416,8 @@ struct GroupedGemmKernel
const auto& d_block_window = gemm_tile_windows.at(Base::I2);
// Get hot-loop and tail configuration
const index_t num_loop =
amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const index_t num_loop = __builtin_amdgcn_readfirstlane(
TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const TailNumber tail_num = GemmPipeline::GetBlockLoopTailNum(num_loop);
// Run GEMM pipeline with compile-time branching

6
include/ck_tile/ops/gemm/kernel/streamk_gemm_kernel.hpp
View File

@@ -271,8 +271,8 @@ struct StreamKKernel
uint32_t block_idx = ck_tile::get_block_1d_id();
bool is_padding_block =
amd_wave_read_first_lane(block_idx >= kargs.tile_partitioner.sk_num_blocks &&
block_idx < kargs.tile_partitioner.dp_start_block_idx);
__builtin_amdgcn_readfirstlane(block_idx >= kargs.tile_partitioner.sk_num_blocks &&
block_idx < kargs.tile_partitioner.dp_start_block_idx);
// Padding blocks make it such that the DP blocks are aligned with the number of CUs; they
// should not partake in the GEMM
@@ -289,7 +289,7 @@ struct StreamKKernel
{
// Determine the number of macro tiles in A and B this WG is resposible for in the
// current C macro tile.
uint32_t current_iter_length = amd_wave_read_first_lane(
uint32_t current_iter_length = __builtin_amdgcn_readfirstlane(
kargs.tile_partitioner.GetCurrentIterLength(iter_start, iter_end));
// Determine the 1D tile_idx and the iter_offset for this WG.

46
include/ck_tile/ops/gemm/kernel/universal_gemm_kernel.hpp
View File

@@ -326,19 +326,19 @@ struct UniversalGemmKernel
__device__ SplitKBatchOffset(const KernelArgs& kargs, const std::size_t k_id = blockIdx.z)
{
constexpr auto K1 = TilePartitioner::BlockGemmShape::WarpTile::at(number<2>{});
const index_t K_t = amd_wave_read_first_lane(kargs.k_batch * K1);
const index_t KRead = amd_wave_read_first_lane((kargs.K + K_t - 1) / K_t * K1);
const index_t K_t = __builtin_amdgcn_readfirstlane(kargs.k_batch * K1);
const index_t KRead = __builtin_amdgcn_readfirstlane((kargs.K + K_t - 1) / K_t * K1);
static_for<0, NumATensor, 1>{}([&](auto index) {
using AiLayout = remove_cvref_t<std::tuple_element_t<index.value, AsLayout>>;
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, AiLayout>)
{
as_k_split_offset[index] = amd_wave_read_first_lane(k_id * KRead);
as_k_split_offset[index] = __builtin_amdgcn_readfirstlane(k_id * KRead);
}
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, AiLayout>)
{
as_k_split_offset[index] =
amd_wave_read_first_lane(k_id * KRead * kargs.stride_As[index]);
__builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_As[index]);
}
});
@@ -347,21 +347,21 @@ struct UniversalGemmKernel
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, BiLayout>)
{
bs_k_split_offset[index] =
amd_wave_read_first_lane(k_id * KRead * kargs.stride_Bs[index]);
__builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_Bs[index]);
}
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, BiLayout>)
{
bs_k_split_offset[index] = amd_wave_read_first_lane(k_id * KRead);
bs_k_split_offset[index] = __builtin_amdgcn_readfirstlane(k_id * KRead);
}
});
if(k_id < static_cast<uint32_t>(kargs.k_batch - 1))
{
splitted_k = amd_wave_read_first_lane(KRead);
splitted_k = __builtin_amdgcn_readfirstlane(KRead);
}
else
{
splitted_k = amd_wave_read_first_lane(kargs.K - KRead * (kargs.k_batch - 1));
splitted_k = __builtin_amdgcn_readfirstlane(kargs.K - KRead * (kargs.k_batch - 1));
}
}
@@ -970,8 +970,8 @@ struct UniversalGemmKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop =
amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const index_t num_loop = __builtin_amdgcn_readfirstlane(
TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
// Run GEMM cooperatively by whole workgroup.
const auto& as_block_window = gemm_tile_windows.at(I0);
@@ -1026,8 +1026,8 @@ struct UniversalGemmKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop =
amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const index_t num_loop = __builtin_amdgcn_readfirstlane(
TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
// Run GEMM cooperatively by whole workgroup.
const auto& as_block_window = gemm_tile_windows.at(I0);
@@ -1052,10 +1052,10 @@ struct UniversalGemmKernel
template <bool U = !PersistentKernel, typename = std::enable_if_t<U>>
CK_TILE_DEVICE void operator()(KernelArgs kargs) const
{
const auto blockId = amd_wave_read_first_lane(blockIdx.x);
const auto blockId = __builtin_amdgcn_readfirstlane(blockIdx.x);
const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockId);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const SplitKBatchOffset splitk_batch_offset(kargs);
@@ -1126,22 +1126,22 @@ struct UniversalGemmKernel
template <bool U = PersistentKernel, typename = std::enable_if_t<U>, typename = void>
CK_TILE_DEVICE void operator()(KernelArgs kargs) const
{
const auto grid_size = amd_wave_read_first_lane(get_grid_size());
const auto grid_size = __builtin_amdgcn_readfirstlane(get_grid_size());
const auto num_tiles =
amd_wave_read_first_lane(TilePartitioner::GridSize(kargs.M, kargs.N));
const auto num_work = amd_wave_read_first_lane(num_tiles * kargs.k_batch);
auto block_id = amd_wave_read_first_lane(get_block_id());
__builtin_amdgcn_readfirstlane(TilePartitioner::GridSize(kargs.M, kargs.N));
const auto num_work = __builtin_amdgcn_readfirstlane(num_tiles * kargs.k_batch);
auto block_id = __builtin_amdgcn_readfirstlane(get_block_id());
while(block_id < num_work)
{
// Get the tile index for this block
const auto tile_idx = amd_wave_read_first_lane(block_id % num_tiles);
const auto tile_idx = __builtin_amdgcn_readfirstlane(block_id % num_tiles);
const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(tile_idx);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
// Get the SplitK offset for this block
const auto k_batch = amd_wave_read_first_lane(block_id / num_tiles);
const auto k_batch = __builtin_amdgcn_readfirstlane(block_id / num_tiles);
const SplitKBatchOffset splitk_batch_offset(kargs, k_batch);
std::array<const ADataType*, NumATensor> as_ptr;

2
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4.hpp
View File

@@ -487,7 +487,7 @@ struct GemmPipelineAgBgCrCompV4 : public BaseGemmPipelineAgBgCrCompV4<Problem>
if(HasHotLoop)
{
// minus 2 because we have ping-pong double buffer.
index_t iCounter = amd_wave_read_first_lane(num_loop - 2);
index_t iCounter = __builtin_amdgcn_readfirstlane(num_loop - 2);
do
{
// ping

4
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v5.hpp
View File

@@ -178,7 +178,7 @@ struct GemmPipelineAgBgCrCompV5 : public BaseGemmPipelineAgBgCrCompV5<Problem>
index_t warp_id = get_warp_id();
index_t operation_id =
amd_wave_read_first_lane(get_warp_id()); // 0 - Memory read, 1 - block-gemm
__builtin_amdgcn_readfirstlane(get_warp_id()); // 0 - Memory read, 1 - block-gemm
auto a_offset = (warp_id == 0) ? make_array(0, 0) : make_array(0, KPerBlock);
auto b_offset = (warp_id == 0) ? make_array(0, 0) : make_array(0, KPerBlock);
@@ -336,7 +336,7 @@ struct GemmPipelineAgBgCrCompV5 : public BaseGemmPipelineAgBgCrCompV5<Problem>
MemoryOpsStep(warp_id);
}
index_t num_compute_steps = amd_wave_read_first_lane(num_loop);
index_t num_compute_steps = __builtin_amdgcn_readfirstlane(num_loop);
while(num_compute_steps > 1)
{
block_sync_lds();

26
include/ck_tile/ops/gemm_quant/kernel/gemm_quant_kernel.hpp
View File

@@ -270,34 +270,34 @@ struct QuantGemmKernel
const std::size_t k_id = blockIdx.z)
{
constexpr auto K1 = TilePartitioner::BlockGemmShape::WarpTile::at(I2);
const index_t K_t = amd_wave_read_first_lane(kargs.k_batch * K1);
const index_t KRead = amd_wave_read_first_lane((kargs.K + K_t - 1) / K_t * K1);
const index_t K_t = __builtin_amdgcn_readfirstlane(kargs.k_batch * K1);
const index_t KRead = __builtin_amdgcn_readfirstlane((kargs.K + K_t - 1) / K_t * K1);
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
a_k_split_offset = amd_wave_read_first_lane(k_id * KRead);
a_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
}
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
a_k_split_offset = amd_wave_read_first_lane(k_id * KRead * kargs.stride_A);
a_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_A);
}
if constexpr(std::is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
{
b_k_split_offset = amd_wave_read_first_lane(k_id * KRead * kargs.stride_B);
b_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead * kargs.stride_B);
}
else if constexpr(std::is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
{
b_k_split_offset = amd_wave_read_first_lane(k_id * KRead);
b_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
}
if(k_id < static_cast<uint32_t>(kargs.k_batch - 1))
{
splitted_k = amd_wave_read_first_lane(KRead);
splitted_k = __builtin_amdgcn_readfirstlane(KRead);
}
else
{
splitted_k = amd_wave_read_first_lane(kargs.K - KRead * (kargs.k_batch - 1));
splitted_k = __builtin_amdgcn_readfirstlane(kargs.K - KRead * (kargs.k_batch - 1));
}
}
@@ -918,8 +918,8 @@ struct QuantGemmKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop =
amd_wave_read_first_lane(TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
const index_t num_loop = __builtin_amdgcn_readfirstlane(
TilePartitioner::GetLoopNum(splitk_batch_offset.splitted_k));
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);
@@ -981,10 +981,10 @@ struct QuantGemmKernel
CK_TILE_DEVICE void operator()(QuantGemmKernelArgs kargs) const
{
const auto blockId = amd_wave_read_first_lane(blockIdx.x);
const auto blockId = __builtin_amdgcn_readfirstlane(blockIdx.x);
const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(blockId);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const SplitKBatchOffset splitk_batch_offset(kargs);
// options

4
include/ck_tile/ops/gemm_quant/kernel/grouped_gemm_quant_kernel.hpp
View File

@@ -305,8 +305,8 @@ struct QuantGroupedGemmKernel
{
const auto [iM, iN] = block_idx_2d;
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const typename Base::SplitKBatchOffset splitk_batch_offset(kargs, block_idx_z);

18
include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_backward_data_kernel.hpp
View File

@@ -840,7 +840,7 @@ struct GroupedConvolutionBackwardDataKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = amd_wave_read_first_lane(TilePartitioner::GetLoopNum(
const index_t num_loop = __builtin_amdgcn_readfirstlane(TilePartitioner::GetLoopNum(
gemm_pad_views.at(I0).get_tensor_descriptor().get_length(I1)));
// Run GEMM cooperatively by whole workgroup.
@@ -891,7 +891,7 @@ struct GroupedConvolutionBackwardDataKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = amd_wave_read_first_lane(
const index_t num_loop = __builtin_amdgcn_readfirstlane(
TilePartitioner::GetLoopNum(gemm_tile_windows.at(I0).get_length(I1)));
// Run GEMM cooperatively by whole workgroup.
@@ -936,7 +936,7 @@ struct GroupedConvolutionBackwardDataKernel
CK_TILE_DEVICE void operator()(GroupedConvBwdDataKernelArgsSpecialized kargs) const
{
const auto blockIdX = amd_wave_read_first_lane(blockIdx.x);
const auto blockIdX = __builtin_amdgcn_readfirstlane(blockIdx.x);
const index_t group_id = FindGroupId(kargs, blockIdX);
const auto [iM, iN] = OffsettedTile1DPartitioner<TilePartitioner>::GetOffsetedTileIndex(
@@ -944,13 +944,13 @@ struct GroupedConvolutionBackwardDataKernel
kargs.c_grid_descs_m_n[group_id].get_length(I0),
kargs.c_grid_descs_m_n[group_id].get_length(I1));
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const auto blockIdY = amd_wave_read_first_lane(blockIdx.y);
const auto group_offset_a = amd_wave_read_first_lane(kargs.group_stride_a * blockIdY);
const auto group_offset_b = amd_wave_read_first_lane(kargs.group_stride_b * blockIdY);
const auto group_offset_c = amd_wave_read_first_lane(kargs.group_stride_c * blockIdY);
const auto blockIdY = __builtin_amdgcn_readfirstlane(blockIdx.y);
const auto group_offset_a = __builtin_amdgcn_readfirstlane(kargs.group_stride_a * blockIdY);
const auto group_offset_b = __builtin_amdgcn_readfirstlane(kargs.group_stride_b * blockIdY);
const auto group_offset_c = __builtin_amdgcn_readfirstlane(kargs.group_stride_c * blockIdY);
// options
// conv_bwd_data = Out * Weight = In

36
include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_backward_weight_kernel.hpp
View File

@@ -423,20 +423,22 @@ struct GroupedConvolutionBackwardWeightKernel
__device__ SplitKBatchOffset(const GroupedConvBwdWeightKernelArgsSpecialized& kargs,
const std::size_t k_id = blockIdx.z)
{
constexpr auto K1 = TilePartitioner::BlockGemmShape::WarpTile::at(number<2>{});
const index_t K_t = amd_wave_read_first_lane(kargs.k_batch * K1);
const index_t KRead = amd_wave_read_first_lane((kargs.GemmK + K_t - 1) / K_t * K1);
constexpr auto K1 = TilePartitioner::BlockGemmShape::WarpTile::at(number<2>{});
const index_t K_t = __builtin_amdgcn_readfirstlane(kargs.k_batch * K1);
const index_t KRead =
__builtin_amdgcn_readfirstlane((kargs.GemmK + K_t - 1) / K_t * K1);
a_k_split_offset = amd_wave_read_first_lane(k_id * KRead);
b_k_split_offset = amd_wave_read_first_lane(k_id * KRead);
a_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
b_k_split_offset = __builtin_amdgcn_readfirstlane(k_id * KRead);
if(k_id < static_cast<uint32_t>(kargs.k_batch - 1))
{
splitted_k = amd_wave_read_first_lane(KRead);
splitted_k = __builtin_amdgcn_readfirstlane(KRead);
}
else
{
splitted_k = amd_wave_read_first_lane(kargs.GemmK - KRead * (kargs.k_batch - 1));
splitted_k =
__builtin_amdgcn_readfirstlane(kargs.GemmK - KRead * (kargs.k_batch - 1));
}
}
@@ -803,22 +805,22 @@ struct GroupedConvolutionBackwardWeightKernel
CK_TILE_DEVICE void operator()(GroupedConvBwdWeightKernelArgsSpecialized kargs) const
{
const auto blockIdX = amd_wave_read_first_lane(blockIdx.x);
const auto blockIdX = __builtin_amdgcn_readfirstlane(blockIdx.x);
const auto [iM, iN] =
TilePartitioner{kargs.GemmM, kargs.GemmN}.GetOutputTileIndex(blockIdX);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const auto blockIdZ = amd_wave_read_first_lane(blockIdx.z);
const index_t num_loop = amd_wave_read_first_lane(
const auto blockIdZ = __builtin_amdgcn_readfirstlane(blockIdx.z);
const index_t num_loop = __builtin_amdgcn_readfirstlane(
ck_tile::integer_divide_ceil(kargs.GemmK, kargs.k_batch * TilePartitioner::KPerBlock));
const index_t i_k =
amd_wave_read_first_lane(blockIdZ * num_loop * TilePartitioner::KPerBlock);
__builtin_amdgcn_readfirstlane(blockIdZ * num_loop * TilePartitioner::KPerBlock);
const auto blockIdY = amd_wave_read_first_lane(blockIdx.y);
const auto group_offset_a = amd_wave_read_first_lane(kargs.group_stride_a * blockIdY);
const auto group_offset_b = amd_wave_read_first_lane(kargs.group_stride_b * blockIdY);
const auto group_offset_c = amd_wave_read_first_lane(kargs.group_stride_c * blockIdY);
const auto blockIdY = __builtin_amdgcn_readfirstlane(blockIdx.y);
const auto group_offset_a = __builtin_amdgcn_readfirstlane(kargs.group_stride_a * blockIdY);
const auto group_offset_b = __builtin_amdgcn_readfirstlane(kargs.group_stride_b * blockIdY);
const auto group_offset_c = __builtin_amdgcn_readfirstlane(kargs.group_stride_c * blockIdY);
// options
// conv_bwd_weight = Out * In = Weight

24
include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_forward_kernel.hpp
View File

@@ -752,7 +752,8 @@ struct GroupedConvolutionForwardKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = amd_wave_read_first_lane(TilePartitioner::GetLoopNum(kargs.GemmK));
const index_t num_loop =
__builtin_amdgcn_readfirstlane(TilePartitioner::GetLoopNum(kargs.GemmK));
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);
@@ -801,7 +802,8 @@ struct GroupedConvolutionForwardKernel
const auto& gemm_pad_views = MakeGemmPadViews(gemm_tensor_views_tuple);
auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
const index_t num_loop = amd_wave_read_first_lane(TilePartitioner::GetLoopNum(kargs.GemmK));
const index_t num_loop =
__builtin_amdgcn_readfirstlane(TilePartitioner::GetLoopNum(kargs.GemmK));
// Run GEMM cooperatively by whole workgroup.
const auto& a_block_window = gemm_tile_windows.at(I0);
@@ -820,22 +822,22 @@ struct GroupedConvolutionForwardKernel
CK_TILE_DEVICE void operator()(GroupedConvFwdKernelArgsSpecialized kargs) const
{
const auto blockIdX = amd_wave_read_first_lane(blockIdx.x);
const auto blockIdX = __builtin_amdgcn_readfirstlane(blockIdx.x);
const auto [iM, iN] =
TilePartitioner{kargs.GemmM, kargs.GemmN}.GetOutputTileIndex(blockIdX);
const index_t i_m = amd_wave_read_first_lane(iM * TilePartitioner::MPerBlock);
const index_t i_n = amd_wave_read_first_lane(iN * TilePartitioner::NPerBlock);
const index_t i_m = __builtin_amdgcn_readfirstlane(iM * TilePartitioner::MPerBlock);
const index_t i_n = __builtin_amdgcn_readfirstlane(iN * TilePartitioner::NPerBlock);
const auto blockIdY = amd_wave_read_first_lane(blockIdx.y);
const auto group_offset_a = amd_wave_read_first_lane(kargs.group_stride_a * blockIdY);
const auto group_offset_b = amd_wave_read_first_lane(kargs.group_stride_b * blockIdY);
const auto group_offset_c = amd_wave_read_first_lane(kargs.group_stride_c * blockIdY);
const auto blockIdY = __builtin_amdgcn_readfirstlane(blockIdx.y);
const auto group_offset_a = __builtin_amdgcn_readfirstlane(kargs.group_stride_a * blockIdY);
const auto group_offset_b = __builtin_amdgcn_readfirstlane(kargs.group_stride_b * blockIdY);
const auto group_offset_c = __builtin_amdgcn_readfirstlane(kargs.group_stride_c * blockIdY);
// Split-N handling: Get which split this workgroup handles
const auto blockIdZ = amd_wave_read_first_lane(blockIdx.z);
const auto blockIdZ = __builtin_amdgcn_readfirstlane(blockIdx.z);
// Calculate batch offset for this split
const index_t batch_offset = amd_wave_read_first_lane(blockIdZ * kargs.n_per_split);
const index_t batch_offset = __builtin_amdgcn_readfirstlane(blockIdZ * kargs.n_per_split);
// Calculate memory offsets for this split
const long_index_t input_batch_offset = static_cast<long_index_t>(batch_offset) *

6
include/ck_tile/ops/image_to_column/kernel/image_to_column_kernel.hpp
View File

@@ -175,9 +175,9 @@ struct ImageToColumn
{
const auto [M, K] = CalculateMKDims(kargs);
const index_t iM = amd_wave_read_first_lane(blockIdx.x * kMPerBlock);
const index_t iK = amd_wave_read_first_lane(blockIdx.y * kKPerBlock);
const index_t iBatch = amd_wave_read_first_lane(blockIdx.z);
const index_t iM = __builtin_amdgcn_readfirstlane(blockIdx.x * kMPerBlock);
const index_t iK = __builtin_amdgcn_readfirstlane(blockIdx.y * kKPerBlock);
const index_t iBatch = __builtin_amdgcn_readfirstlane(blockIdx.z);
const auto in_offset = iBatch * kargs.image_g_n_c_wis_strides[I0];
const auto out_offset = iBatch * kargs.gemm_g_m_k_strides[I0];

6
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.hpp
View File

@@ -99,7 +99,7 @@ struct Layernorm2dFwdPipelineTwoPass
// Problem::BlockShape
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
index_t num_n_tile_iteration =
amd_wave_read_first_lane(integer_divide_ceil(row_size, Block_N));
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
// total number of count assume current iter have no pad(only last iter has pad)
constexpr index_t count_per_iter =
@@ -119,7 +119,7 @@ struct Layernorm2dFwdPipelineTwoPass
auto mean = block_norm_reduce.template MakeMeanVarBlockTile<XTensorType>();
auto var = block_norm_reduce.template MakeMeanVarBlockTile<XTensorType>();
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
@@ -197,7 +197,7 @@ struct Layernorm2dFwdPipelineTwoPass
move_tile_window(y_window, {0, stride_to_right_most_window});
// layernorm computation
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto acc = make_static_distributed_tensor<ComputeDataType>(
decltype(load_tile(x_window))::get_tile_distribution());

4
include/ck_tile/ops/reduce/kernel/reduce2d_kernel.hpp
View File

@@ -156,7 +156,7 @@ struct Reduce
const auto merged_reduce_len =
transformed_x_tensor.get_tensor_descriptor().get_lengths().at(number<1>{});
index_t num_n_tile_iteration =
amd_wave_read_first_lane(integer_divide_ceil(merged_reduce_len, S::Block_N));
__builtin_amdgcn_readfirstlane(integer_divide_ceil(merged_reduce_len, S::Block_N));
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
@@ -167,7 +167,7 @@ struct Reduce
auto y_compute = block_reduce2d.template MakeYBlockTile<XTensorType>();
set_tile(y_compute, reduce_func.template GetIdentityValue<ComputeDataType>());
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto x = load_tile(x_window);
block_reduce2d(x, y_compute, reduce_func);

6
include/ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_two_pass.hpp
View File

@@ -82,7 +82,7 @@ struct Rmsnorm2dFwdPipelineTwoPass
// Problem::BlockShape
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
index_t num_n_tile_iteration =
amd_wave_read_first_lane(integer_divide_ceil(row_size, Block_N));
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
auto reduce_square_sum_func = ReduceOp::SquareAdd{};
auto reduce_sum_func = ReduceOp::Add{};
@@ -95,7 +95,7 @@ struct Rmsnorm2dFwdPipelineTwoPass
auto square_sum = block_reduce2d.template MakeYBlockTile<ComputeTensorType>();
set_tile(square_sum, reduce_square_sum_func.GetIdentityValue<ComputeDataType>());
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
@@ -151,7 +151,7 @@ struct Rmsnorm2dFwdPipelineTwoPass
move_tile_window(y_window, {0, stride_to_right_most_window});
// rmsnorm computation
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto acc = make_static_distributed_tensor<ComputeDataType>(
decltype(load_tile(x_window))::get_tile_distribution());

2
include/ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp
View File

@@ -138,7 +138,7 @@ struct MoeSmoothquant
const index_t i_topk = blockIdx.x;
const index_t i_token = blockIdx.y * Block_M;
const index_t i_token_in_thrd =
amd_wave_read_first_lane(threadIdx.x / Problem::BlockShape::ThreadPerBlock_N);
__builtin_amdgcn_readfirstlane(threadIdx.x / Problem::BlockShape::ThreadPerBlock_N);
const index_t i_expert = reinterpret_cast<const index_t*>(
kargs.p_topk_ids)[(i_token + i_token_in_thrd) * kargs.topk + i_topk];

6
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp
View File

@@ -57,7 +57,7 @@ struct SmoothquantPipelineTwoPass
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
index_t num_n_tile_iteration =
amd_wave_read_first_lane(integer_divide_ceil(row_size, Block_N));
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
@@ -77,7 +77,7 @@ struct SmoothquantPipelineTwoPass
auto absmax = block_reduce2d.template MakeYBlockTile<XTensorType>();
set_tile(absmax, reduce_absmax_func.GetIdentityValue<ComputeDataType>());
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto x = load_tile(x_window);
const auto smscale = load_tile(smscale_window);
@@ -121,7 +121,7 @@ struct SmoothquantPipelineTwoPass
move_tile_window(qy_window, {0, stride_to_right_most_window});
// recompute y and quantize y to qy
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto x = load_tile(x_window);
const auto smscale = load_tile(smscale_window);

6
include/ck_tile/ops/topk_softmax/kernel/topk_softmax_kernel.hpp
View File

@@ -96,9 +96,9 @@ struct TopkSoftmaxKernel
if(block_row_id > kargs.num_rows)
return;
index_t block_os_inp = amd_wave_read_first_lane(block_row_id * kargs.stride_input);
index_t block_os_out = amd_wave_read_first_lane(block_row_id * kargs.stride_output);
index_t num_rows_rem = amd_wave_read_first_lane(kargs.num_rows - block_row_id);
index_t block_os_inp = __builtin_amdgcn_readfirstlane(block_row_id * kargs.stride_input);
index_t block_os_out = __builtin_amdgcn_readfirstlane(block_row_id * kargs.stride_output);
index_t num_rows_rem = __builtin_amdgcn_readfirstlane(kargs.num_rows - block_row_id);
const auto input_window = [&]() {
const InputType* p_input =