Skip to content

GitLab

Commit 97ba755f authored by Chao Liu's avatar Chao Liu
Browse files

refactor

parent 8d460740
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 155 additions and 528 deletions
driver/device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp
View file @ 97ba755f
......@@ -38,7 +38,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
// reorder weight
auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
auto wei_cyxk_desc = make_ConstantTensorDescriptor_packed(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
......@@ -51,7 +51,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
std::thread::hardware_concurrency());
// reorder input
auto in_chwn_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Hi, Wi, N>{});
auto in_chwn_desc = make_ConstantTensorDescriptor_packed(Sequence<C, Hi, Wi, N>{});
ostream_ConstantTensorDescriptor(in_chwn_desc, std::cout << "in_chwn_desc: ");
Tensor<T> in_chwn(make_TensorDescriptor(in_chwn_desc));
......@@ -64,8 +64,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
std::thread::hardware_concurrency());
// output
auto out_khwn_desc =
make_ConstantTensorDescriptor_default_rank_packed(Sequence<K, Ho, Wo, N>{});
auto out_khwn_desc = make_ConstantTensorDescriptor_packed(Sequence<K, Ho, Wo, N>{});
ostream_ConstantTensorDescriptor(out_khwn_desc, std::cout << "out_khwn_desc: ");
Tensor<T> out_khwn(make_TensorDescriptor(out_khwn_desc));
......
driver/device_convolution_implicit_gemm_v1_nchw_cyxk_khwn.hpp
View file @ 97ba755f
......@@ -37,7 +37,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
// reorder weight
auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
auto wei_cyxk_desc = make_ConstantTensorDescriptor_packed(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
......@@ -50,8 +50,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
std::thread::hardware_concurrency());
// output
auto out_khwn_desc =
make_ConstantTensorDescriptor_default_rank_packed(Sequence<K, Ho, Wo, N>{});
auto out_khwn_desc = make_ConstantTensorDescriptor_packed(Sequence<K, Ho, Wo, N>{});
ostream_ConstantTensorDescriptor(out_khwn_desc, std::cout << "out_khwn_desc: ");
Tensor<T> out_khwn(make_TensorDescriptor(out_khwn_desc));
......
driver/device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw.hpp
View file @ 97ba755f
......@@ -36,7 +36,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
// reorder weight
auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
auto wei_cyxk_desc = make_ConstantTensorDescriptor_packed(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
......
driver/device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp
View file @ 97ba755f
......@@ -36,7 +36,7 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
// reorder weight
auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
auto wei_cyxk_desc = make_ConstantTensorDescriptor_packed(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
......
driver/driver.hip.cpp
View file @ 97ba755f
......@@ -443,7 +443,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
#elif 1
// 3x3 filter, 28x28 image
constexpr index_t N = 128;
constexpr index_t C = 256;
......@@ -455,7 +455,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 1
#elif 0
// 1x1 filter, 28x28 image
constexpr index_t N = 128;
constexpr index_t C = 512;
......@@ -568,8 +568,8 @@ int main(int argc, char* argv[])
auto lower_pads = Sequence<HPad, WPad>{};
auto upper_pads = Sequence<HPad, WPad>{};
auto in_nchw_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<N, C, HI, WI>{});
auto wei_kcyx_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<K, C, Y, X>{});
auto in_nchw_desc = make_ConstantTensorDescriptor_packed(Sequence<N, C, HI, WI>{});
auto wei_kcyx_desc = make_ConstantTensorDescriptor_packed(Sequence<K, C, Y, X>{});
auto out_nkhw_desc = get_convolution_with_padding_output_default_4d_tensor_descriptor(
in_nchw_desc, wei_kcyx_desc, lower_pads, upper_pads);
......
src/include/ConstantMergedTensorDescriptor.hip.hpp
View file @ 97ba755f
......@@ -114,7 +114,7 @@ struct ConstantMergedTensorDescriptor
__host__ __device__ static Array<index_t, nDim> GetMultiIndexFrom1dIndex(index_t id)
{
constexpr auto dummy_desc = make_ConstantTensorDescriptor_default_rank_packed(GetLengths());
constexpr auto dummy_desc = make_ConstantTensorDescriptor_packed(GetLengths());
return dummy_desc.GetMultiIndexFrom1dIndex(id);
}
......@@ -128,7 +128,7 @@ __host__ __device__ constexpr auto make_ConstantMergedTensorDescriptor(OriginalT
}
template <class TDesc>
__host__ __device__ void print_ConstantMergedTensorDescriptor(TDesc, const char* s)
__host__ __device__ void print_ConstantMergedTensorDescriptor(const char* s, TDesc)
{
print_ConstantTensorDescriptor(TDesc::GetOriginalTensorDescriptor(), s);
print_ConstantTensorDescriptor(s, TDesc::GetOriginalTensorDescriptor());
}
src/include/blockwise_4d_tensor_op.hip.hpp
View file @ 97ba755f
......@@ -13,7 +13,7 @@ blockwise_4d_tensor_pointwise_operation_unary(DstDesc, Float* __restrict__ p_dst
constexpr auto dst_desc = DstDesc{};
constexpr auto desc = make_ConstantTensorDescriptor_default_rank_packed(dst_desc.GetLengths());
constexpr auto desc = make_ConstantTensorDescriptor_packed(dst_desc.GetLengths());
#if 0
if(get_thread_local_1d_id() == 0)
......@@ -108,7 +108,7 @@ __device__ void blockwise_4d_tensor_pointwise_operation_binary_reorder_by_get_ds
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr auto ref_desc = make_ConstantTensorDescriptor_default_rank_packed(SrcOpLengths{});
constexpr auto ref_desc = make_ConstantTensorDescriptor_packed(SrcOpLengths{});
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
......@@ -259,7 +259,7 @@ struct Blockwise4dTensorCopy1
constexpr index_t read_per_d3 = mod_conv::integer_divide_ceil(L3, DataPerRead);
constexpr auto ref_desc =
make_ConstantTensorDescriptor_default_rank_packed(Sequence<L0, L1, L2, read_per_d3>{});
make_ConstantTensorDescriptor_packed(Sequence<L0, L1, L2, read_per_d3>{});
constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
......@@ -336,7 +336,7 @@ struct BlockwiseChwnTensorCopyPadded
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr auto ref_desc = make_ConstantTensorDescriptor_default_rank_packed(DstOpLengths{});
constexpr auto ref_desc = make_ConstantTensorDescriptor_packed(DstOpLengths{});
constexpr auto h_global_pad_low = GlobalLowerPads{}.Get(I0);
constexpr auto w_global_pad_low = GlobalLowerPads{}.Get(I1);
......@@ -510,8 +510,7 @@ struct Blockwise4dTensorCopy3
}
}
constexpr auto thread_cluster_desc =
make_ConstantTensorDescriptor_default_rank_packed(ThreadPerDims{});
constexpr auto thread_cluster_desc = make_ConstantTensorDescriptor_packed(ThreadPerDims{});
const auto thread_multi_id =
thread_cluster_desc.GetMultiIndexFrom1dIndex(get_thread_local_1d_id());
......@@ -653,7 +652,7 @@ struct Blockwise4dTensorCopy3
constexpr index_t nloop_d2 = L2 / thread_per_d2;
constexpr index_t nloop_d3 = mod_conv::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
constexpr auto clipboard_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto clipboard_desc = make_ConstantTensorDescriptor_packed(
Sequence<nloop_d0, nloop_d1, nloop_d2, nloop_d3 * DataPerRead>{});
#pragma unroll
......@@ -720,7 +719,7 @@ struct Blockwise4dTensorCopy3
constexpr index_t nloop_d2 = L2 / thread_per_d2;
constexpr index_t nloop_d3 = mod_conv::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
constexpr auto clipboard_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto clipboard_desc = make_ConstantTensorDescriptor_packed(
Sequence<nloop_d0, nloop_d1, nloop_d2, nloop_d3 * DataPerRead>{});
#pragma unroll
......
src/include/blockwise_generic_tensor_slice_op.hip.hpp
View file @ 97ba755f
......@@ -63,7 +63,7 @@ struct BlockwiseGenericTensorSliceCopy_v1
"wrong!");
// thread cluster
constexpr auto thread_cluster_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto thread_cluster_desc = make_ConstantTensorDescriptor_packed(
DataClusterLengths{}.ReorderGivenNew2Old(ThreadClusterArrangeOrder{}));
// BlockSize
......@@ -185,7 +185,7 @@ struct BlockwiseGenericTensorSliceCopy_v1
constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * DataClusterLengths{});
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_default_rank_packed(SubLengths{} * repeat_lengths);
make_ConstantTensorDescriptor_packed(SubLengths{} * repeat_lengths);
return thread_tensor_desc.GetElementSpace();
}
......@@ -199,8 +199,8 @@ struct BlockwiseGenericTensorSliceCopy_v1
constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * DataClusterLengths{});
constexpr auto thread_tensor_desc = make_ConstantTensorDescriptor_default_rank_packed(
thread_sub_tensor_lengths * repeat_lengths);
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_packed(thread_sub_tensor_lengths * repeat_lengths);
static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
constexpr auto repeat_multi_id = sequence2array(decltype(repeat_multi_id_){});
......@@ -237,8 +237,8 @@ struct BlockwiseGenericTensorSliceCopy_v1
constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * DataClusterLengths{});
constexpr auto thread_tensor_desc = make_ConstantTensorDescriptor_default_rank_packed(
thread_sub_tensor_lengths * repeat_lengths);
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_packed(thread_sub_tensor_lengths * repeat_lengths);
static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
constexpr auto repeat_multi_id = sequence2array(decltype(repeat_multi_id_){});
......
src/include/blockwise_tensor_slice_op.hip.hpp
View file @ 97ba755f
......@@ -40,7 +40,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
src_cluster_lengths.ReorderGivenNew2Old(map_thread_cluster_2_src_cluster);
constexpr auto thread_cluster_desc =
make_ConstantTensorDescriptor_default_rank_packed(thread_cluster_lengths);
make_ConstantTensorDescriptor_packed(thread_cluster_lengths);
// sanity check: data type
static_assert(is_same<Float, float>::value, "wrong! only support float for now!\n");
......@@ -175,7 +175,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_default_rank_packed(thread_tensor_lengths);
make_ConstantTensorDescriptor_packed(thread_tensor_lengths);
return thread_tensor_desc.GetElementSpace();
}
......@@ -196,7 +196,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_default_rank_packed(thread_tensor_lengths);
make_ConstantTensorDescriptor_packed(thread_tensor_lengths);
static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
......@@ -234,7 +234,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
constexpr auto thread_tensor_desc =
make_ConstantTensorDescriptor_default_rank_packed(thread_tensor_lengths);
make_ConstantTensorDescriptor_packed(thread_tensor_lengths);
static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
......
src/include/conv_common.hip.hpp
View file @ 97ba755f
......@@ -30,7 +30,7 @@ __host__ __device__ constexpr auto get_convolution_output_default_4d_tensor_desc
constexpr auto HO = HI + 1 - Y;
constexpr auto WO = WI + 1 - X;
return make_ConstantTensorDescriptor_default_rank_packed(Sequence<N, K, HO, WO>{});
return make_ConstantTensorDescriptor_packed(Sequence<N, K, HO, WO>{});
}
template <class InDesc, class WeiDesc, class LowerPads, class UpperPads>
......@@ -67,7 +67,7 @@ __host__ __device__ constexpr auto get_convolution_with_padding_output_default_4
constexpr auto HO = HI + HPadLow + HPadUp + 1 - Y;
constexpr auto WO = WI + WPadLow + WPadUp + 1 - X;
return make_ConstantTensorDescriptor_default_rank_packed(Sequence<N, K, HO, WO>{});
return make_ConstantTensorDescriptor_packed(Sequence<N, K, HO, WO>{});
}
template <class InDesc, class WeiDesc, class OutDesc>
......
src/include/gridwise_convolution_direct_v2_nchw_kcyx_nkhw.hip.hpp
View file @ 97ba755f
......@@ -45,23 +45,23 @@ struct GridwiseConvolutionDirect_v2_nchw_kcyx_nkhw
constexpr index_t Y = wei_kcyx_global_desc.GetLength(I2);
constexpr index_t X = wei_kcyx_global_desc.GetLength(I3);
constexpr auto wei_ke_global_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto wei_ke_global_desc = make_ConstantTensorDescriptor_packed(
Sequence<K, C * Y * X>{}); // 2d view of wei for blockwise copy
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
constexpr auto in_nchw_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto in_nchw_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<NPerBlock, CPerBlock, HiPerBlock, WiPerBlock>{},
Number<InBlockCopyDataPerRead>{});
constexpr auto wei_ke_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_ke_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<KPerBlock, CPerBlock * Y * X>{},
Number<WeiBlockCopyDataPerRead>{}); // 2d view of wei for blockwise copy
constexpr auto wei_kcyx_block_desc = make_ConstantTensorDescriptor_default_rank(
Sequence<KPerBlock, CPerBlock, Y, X>{},
Sequence<wei_ke_block_desc.GetStride(I0), Y * X, X, 1>{});
constexpr auto wei_kcyx_block_desc =
make_ConstantTensorDescriptor(Sequence<KPerBlock, CPerBlock, Y, X>{},
Sequence<wei_ke_block_desc.GetStride(I0), Y * X, X, 1>{});
// shared mem
constexpr index_t in_block_element_size =
......@@ -82,11 +82,11 @@ struct GridwiseConvolutionDirect_v2_nchw_kcyx_nkhw
constexpr index_t HiPerThread = HoPerThread + Y - 1;
constexpr index_t WiPerThread = WoPerThread + X - 1;
constexpr auto in_nchw_thread_block_desc = make_ConstantTensorDescriptor_default_rank(
constexpr auto in_nchw_thread_block_desc = make_ConstantTensorDescriptor(
Sequence<NPerThread, CPerThread, HiPerThread, WiPerThread>{},
in_nchw_block_desc.GetStrides());
constexpr auto wei_kcyx_thread_block_desc = make_ConstantTensorDescriptor_default_rank(
constexpr auto wei_kcyx_thread_block_desc = make_ConstantTensorDescriptor(
Sequence<KPerThread, CPerThread, Y, X>{}, wei_kcyx_block_desc.GetStrides());
constexpr auto out_nkhw_thread_desc = get_convolution_output_default_4d_tensor_descriptor(
......
src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_chwn_cyxk_khwn.hip.hpp
View file @ 97ba755f
......@@ -85,7 +85,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
constexpr index_t WBlockWork = mod_conv::integer_divide_ceil(Wo, WoPerBlock);
constexpr index_t NBlockWork = mod_conv::integer_divide_ceil(N, NPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto block_work_desc = make_ConstantTensorDescriptor_packed(
Sequence<KBlockWork, HBlockWork, WBlockWork, NBlockWork>{});
const auto block_work_multi_id =
......@@ -109,7 +109,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockCopyDataPerRead_N>{});
......@@ -118,12 +118,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
......
src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_khwn.hip.hpp
View file @ 97ba755f
......@@ -86,7 +86,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
constexpr index_t HBlockWork = mod_conv::integer_divide_ceil(Ho, HoPerBlock);
constexpr index_t WBlockWork = mod_conv::integer_divide_ceil(Wo, WoPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto block_work_desc = make_ConstantTensorDescriptor_packed(
Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
const auto block_work_multi_id =
......@@ -102,7 +102,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
// global tensor view
constexpr auto wei_c_k_global_desc =
make_ConstantTensorDescriptor_default_rank(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
make_ConstantTensorDescriptor(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
// LDS tensor view
// be careful of alignment
......@@ -111,7 +111,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockReorderDataPerWrite_N>{});
......@@ -120,12 +120,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
......@@ -448,10 +448,10 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
constexpr index_t K1 = KPerBlock / KPerThread;
#if 0
constexpr auto out_10d_global_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto out_10d_global_desc = make_ConstantTensorDescriptor_packed(
Sequence<K / (K1 * K2), K1, K2, Ho, Wo / (W1 * W2 * W3), W1, W2, W3, N / N1, N1>{});
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, W3, 1, N1>{});
#else
constexpr auto out_10d_global_desc =
......
src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw.hip.hpp
View file @ 97ba755f
......@@ -86,7 +86,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw
constexpr index_t HBlockWork = mod_conv::integer_divide_ceil(Ho, HoPerBlock);
constexpr index_t WBlockWork = mod_conv::integer_divide_ceil(Wo, WoPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto block_work_desc = make_ConstantTensorDescriptor_packed(
Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
const auto block_work_multi_id =
......@@ -110,7 +110,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockReorderDataPerWrite_N>{});
......@@ -119,12 +119,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
......
src/include/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw.hip.hpp
View file @ 97ba755f
......@@ -83,7 +83,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
constexpr index_t HBlockWork = mod_conv::integer_divide_ceil(Ho, HoPerBlock);
constexpr index_t WBlockWork = mod_conv::integer_divide_ceil(Wo, WoPerBlock);
constexpr auto block_work_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto block_work_desc = make_ConstantTensorDescriptor_packed(
Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
const auto block_work_multi_id =
......@@ -99,7 +99,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
// global tensor view
constexpr auto wei_c_k_global_desc =
make_ConstantTensorDescriptor_default_rank(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
make_ConstantTensorDescriptor(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
// LDS tensor view
// be careful of alignment
......@@ -108,7 +108,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
GemmDataPerReadA,
GemmDataPerReadB);
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
Number<InBlockReorderDataPerWrite_N>{});
......@@ -117,12 +117,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
"GemmDataPerReadB alignment requirement is not meet");
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
// tensor view of threadwise output in register
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
// blockwise copy
......
src/include/gridwise_convolution_implicit_gemm_v3_lds_double_buffer_nchw_cyxk_nkhw.hip.hpp
View file @ 97ba755f
......@@ -88,7 +88,7 @@ struct GridwiseConvolutionImplicitGemm_v3_lds_double_buffer_nchw_cyxk_nkhw
constexpr index_t BBlockWork = B / BPerBlock;
constexpr auto block_work_desc =
make_ConstantTensorDescriptor_default_rank_packed(Sequence<KBlockWork, BBlockWork>{});
make_ConstantTensorDescriptor_packed(Sequence<KBlockWork, BBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
......@@ -111,9 +111,8 @@ struct GridwiseConvolutionImplicitGemm_v3_lds_double_buffer_nchw_cyxk_nkhw
// memory layout descriptor in LDS [C, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_c_n1_b_n2_block_mem_desc =
make_ConstantTensorDescriptor_default_rank_aligned(
Sequence<CPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
constexpr auto in_c_n1_b_n2_block_mem_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
......@@ -143,7 +142,7 @@ struct GridwiseConvolutionImplicitGemm_v3_lds_double_buffer_nchw_cyxk_nkhw
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDataPerAccess_K, GemmDataPerReadA)>{});
......@@ -367,7 +366,7 @@ struct GridwiseConvolutionImplicitGemm_v3_lds_double_buffer_nchw_cyxk_nkhw
// define tensor descriptor for threadwise copy
// output memory layout descriptor in register
constexpr auto out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc =
make_ConstantTensorDescriptor_default_rank_packed(
make_ConstantTensorDescriptor_packed(
Sequence<KPerBlock / (K1 * K2), 1, K2, N1, 1, 1, 1, N2>{});
// output tensor descriptor in register, src of threadwise copy
......
src/include/gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hip.hpp
View file @ 97ba755f
......@@ -91,7 +91,7 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
constexpr index_t BBlockWork = B / BPerBlock;
constexpr auto block_work_desc =
make_ConstantTensorDescriptor_default_rank_packed(Sequence<KBlockWork, BBlockWork>{});
make_ConstantTensorDescriptor_packed(Sequence<KBlockWork, BBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
......@@ -114,9 +114,8 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
// memory layout descriptor in LDS [C, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_c_n1_b_n2_block_mem_desc =
make_ConstantTensorDescriptor_default_rank_aligned(
Sequence<CPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
constexpr auto in_c_n1_b_n2_block_mem_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
// TODO: need to properly implement tensor descriptor with alignment
......@@ -146,7 +145,7 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDataPerAccess_K, GemmDataPerReadA)>{});
......@@ -320,7 +319,7 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
// define tensor descriptor for threadwise copy
// output memory layout descriptor in register
constexpr auto out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc =
make_ConstantTensorDescriptor_default_rank_packed(
make_ConstantTensorDescriptor_packed(
Sequence<KPerBlock / (K1 * K2), 1, K2, N1, 1, 1, 1, N2>{});
// output tensor descriptor in register, src of threadwise copy
......
src/include/gridwise_convolution_implicit_gemm_v4_lds_double_buffer_nchw_kcyx_nkhw.hip.hpp
View file @ 97ba755f
......@@ -99,7 +99,7 @@ struct GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
constexpr index_t BBlockWork = B / BPerBlock;
constexpr auto block_work_desc =
make_ConstantTensorDescriptor_default_rank_packed(Sequence<KBlockWork, BBlockWork>{});
make_ConstantTensorDescriptor_packed(Sequence<KBlockWork, BBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
......@@ -127,20 +127,9 @@ struct GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
Sequence<3, 6, 7>{},
Sequence<5>{});
#if 0
if(get_block_1d_id() == 0 && get_thread_local_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_n0_n1_n2_h_w_global_desc,
"in_n0_n1_n2_h_w_global_desc: ");
print_ConstantTensorDescriptor(in_c_y_x_global_desc, "in_c_y_x_global_desc: ");
print_ConstantMergedTensorDescriptor(in_e_n1_b_n2_global_merged_desc,
"in_e_n1_b_n2_global_merged_desc: ");
}
#endif
// memory layout descriptor in LDS [E, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_e_n1_b_n2_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto in_e_n1_b_n2_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
......@@ -174,7 +163,7 @@ struct GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDstDataPerWrite_K, GemmDataPerReadA)>{});
......@@ -406,7 +395,7 @@ struct GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
// define tensor descriptor for threadwise copy
// output memory layout descriptor in register
constexpr auto out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc =
make_ConstantTensorDescriptor_default_rank_packed(
make_ConstantTensorDescriptor_packed(
Sequence<KPerBlock / (K1 * K2), 1, K2, N1, 1, 1, 1, N2>{});
// output tensor descriptor in register, src of threadwise copy
......
src/include/gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hip.hpp
View file @ 97ba755f
......@@ -93,7 +93,7 @@ struct GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
constexpr index_t BBlockWork = B / BPerBlock;
constexpr auto block_work_desc =
make_ConstantTensorDescriptor_default_rank_packed(Sequence<KBlockWork, BBlockWork>{});
make_ConstantTensorDescriptor_packed(Sequence<KBlockWork, BBlockWork>{});
const auto block_work_multi_id =
block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
......@@ -134,7 +134,7 @@ struct GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
// memory layout descriptor in LDS [E, N1, B, N2], dst of blockwise copy
// be careful of LDS alignment
constexpr auto in_e_n1_b_n2_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto in_e_n1_b_n2_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
// this check is ad-hoc
......@@ -167,7 +167,7 @@ struct GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
// tensor descriptor in LDS, dst of blockwise copy
// be careful of LDS alignment
constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<EPerBlock, KPerBlock>{},
Number<mod_conv::max(WeiBlockCopyDstDataPerWrite_K, GemmDataPerReadA)>{});
......@@ -288,7 +288,7 @@ struct GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
// define tensor descriptor for threadwise copy
// output memory layout descriptor in register
constexpr auto out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc =
make_ConstantTensorDescriptor_default_rank_packed(
make_ConstantTensorDescriptor_packed(
Sequence<KPerBlock / (K1 * K2), 1, K2, N1, 1, 1, 1, N2>{});
// output tensor descriptor in register, src of threadwise copy
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment