Skip to content

GitLab

Commit 1a306e0d authored by danyao12's avatar danyao12
Browse files

add OutputDataType&Deterministic for pt1q2

parent 35b59ac6
Hide whitespace changes
Inline Side-by-side
Showing with 367 additions and 274 deletions
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_v4.cpp
View file @ 1a306e0d
...@@ -51,10 +51,11 @@ Kernel outputs: ...@@ -51,10 +51,11 @@ Kernel outputs:
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using F16 = ck::half_t; using F16 = ck::half_t;
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using F32 = float; using F32 = float;
using U16 = unsigned short; using U16 = unsigned short;
using INT32 = int32_t;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale; using Scale = ck::tensor_operation::element_wise::Scale;
...@@ -62,12 +63,13 @@ using Scale = ck::tensor_operation::element_wise::Scale; ...@@ -62,12 +63,13 @@ using Scale = ck::tensor_operation::element_wise::Scale;
using QKVElementOp = PassThrough; using QKVElementOp = PassThrough;
using YElementOp = PassThrough; using YElementOp = PassThrough;
using DataType = F16; using InputDataType = F16;
using OutputDataType = F16;
using GemmDataType = F16; using GemmDataType = F16;
using AccDataType = F32; using AccDataType = F32;
using ShuffleDataType = F32; using ShuffleDataType = F32;
using LSEDataType = F32; using LSEDataType = F32;
using ZDataType = U16; using ZDataType = U16; // INT32
using Acc0BiasDataType = ck::Tuple<>; using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>; using Acc1BiasDataType = ck::Tuple<>;
...@@ -76,6 +78,9 @@ static constexpr ck::index_t NumDimM = 1; ...@@ -76,6 +78,9 @@ static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1; static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1; static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1; static constexpr ck::index_t NumDimO = 1;
// When OutputDataType == F32, CShuffleBlockTransferScalarPerVector_NPerBlock = 4
// When OutputDataType == F16/BF16, CShuffleBlockTransferScalarPerVector_NPerBlock = 8
static constexpr ck::index_t CShuffleBlockTransferScalarPerVector_NPerBlock = 8;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding; static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
#if USING_MASK #if USING_MASK
...@@ -86,10 +91,11 @@ static constexpr auto MaskingSpec = ...@@ -86,10 +91,11 @@ static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled; ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
#endif #endif
static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecialization::Default; static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default; static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default; static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecY = ck::tensor_operation::device::TensorSpecialization::Default; static constexpr auto TensorSpecY = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr bool Deterministic = true;
// DIM should be a multiple of 8. // DIM should be a multiple of 8.
// If DIM <= 32 , ues prototype1 1st template. // If DIM <= 32 , ues prototype1 1st template.
...@@ -103,7 +109,8 @@ using DeviceGemmInstance = ...@@ -103,7 +109,8 @@ using DeviceGemmInstance =
NumDimN, NumDimN,
NumDimK, NumDimK,
NumDimO, NumDimO,
DataType, InputDataType,
OutputDataType,
GemmDataType, GemmDataType,
ZDataType, ZDataType,
LSEDataType, LSEDataType,
...@@ -154,8 +161,9 @@ using DeviceGemmInstance = ...@@ -154,8 +161,9 @@ using DeviceGemmInstance =
1, // CShuffleMXdlPerWavePerShuffle 1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle 1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock CShuffleBlockTransferScalarPerVector_NPerBlock, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec>; // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 64) #elif(DIM <= 64)
using DeviceGemmInstance = using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1< ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1<
...@@ -164,7 +172,8 @@ using DeviceGemmInstance = ...@@ -164,7 +172,8 @@ using DeviceGemmInstance =
NumDimN, NumDimN,
NumDimK, NumDimK,
NumDimO, NumDimO,
DataType, InputDataType,
OutputDataType,
GemmDataType, GemmDataType,
ZDataType, ZDataType,
LSEDataType, LSEDataType,
...@@ -215,8 +224,9 @@ using DeviceGemmInstance = ...@@ -215,8 +224,9 @@ using DeviceGemmInstance =
1, // CShuffleMXdlPerWavePerShuffle 1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle 2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock CShuffleBlockTransferScalarPerVector_NPerBlock, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec>; // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>;
// using DeviceGemmInstance = // using DeviceGemmInstance =
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2< // ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2<
...@@ -225,7 +235,8 @@ using DeviceGemmInstance = ...@@ -225,7 +235,8 @@ using DeviceGemmInstance =
// NumDimN, // NumDimN,
// NumDimK, // NumDimK,
// NumDimO, // NumDimO,
// DataType, // InputDataType,
// OutputDataType,
// GemmDataType, // GemmDataType,
// ZDataType, // ZDataType,
// LSEDataType, // LSEDataType,
...@@ -283,8 +294,9 @@ using DeviceGemmInstance = ...@@ -283,8 +294,9 @@ using DeviceGemmInstance =
// 1, // CShuffleMXdlPerWavePerShuffle // 1, // CShuffleMXdlPerWavePerShuffle
// 2, // CShuffleNXdlPerWavePerShuffle // 2, // CShuffleNXdlPerWavePerShuffle
// S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock // S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
// 8, // CShuffleBlockTransferScalarPerVector_NPerBlock // CShuffleBlockTransferScalarPerVector_NPerBlock,
// MaskingSpec>; // MaskingSpecialization // MaskingSpec,
// Deterministic>;
#elif(DIM <= 128) #elif(DIM <= 128)
using DeviceGemmInstance = using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2< ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2<
...@@ -293,7 +305,8 @@ using DeviceGemmInstance = ...@@ -293,7 +305,8 @@ using DeviceGemmInstance =
NumDimN, NumDimN,
NumDimK, NumDimK,
NumDimO, NumDimO,
DataType, InputDataType,
OutputDataType,
GemmDataType, GemmDataType,
ZDataType, ZDataType,
LSEDataType, LSEDataType,
...@@ -351,14 +364,15 @@ using DeviceGemmInstance = ...@@ -351,14 +364,15 @@ using DeviceGemmInstance =
1, // CShuffleMXdlPerWavePerShuffle 1, // CShuffleMXdlPerWavePerShuffle
4, // CShuffleNXdlPerWavePerShuffle 4, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock CShuffleBlockTransferScalarPerVector_NPerBlock, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec>; // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>;
#endif #endif
// Ref Gemm0: S = alpha * Q * K^T // Ref Gemm0: S = alpha * Q * K^T
// fp16 in, fp32 out // fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<DataType, using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<InputDataType,
DataType, InputDataType,
AccDataType, AccDataType,
AccDataType, AccDataType,
PassThrough, PassThrough,
...@@ -368,13 +382,13 @@ using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm< ...@@ -368,13 +382,13 @@ using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
// Ref Softmax: P = Softmax(S) // Ref Softmax: P = Softmax(S)
// fp32 in, fp16 out // fp32 in, fp16 out
using ReferenceSoftmaxInstance = using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<AccDataType, DataType, AccDataType>; ck::tensor_operation::host::ReferenceSoftmax<AccDataType, InputDataType, AccDataType>;
// Ref Gemm1: Y = P * V // Ref Gemm1: Y = P * V
// fp16 in, fp16 out // fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<DataType, using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<InputDataType,
DataType, InputDataType,
DataType, InputDataType,
AccDataType, AccDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
...@@ -382,16 +396,25 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm< ...@@ -382,16 +396,25 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
// Ref Gemm for backward pass // Ref Gemm for backward pass
// fp16 in, fp16 out // fp16 in, fp16 out
using ReferenceGemmGradInstance = ck::tensor_operation::host::ReferenceBatchedGemm<DataType, using ReferenceGemm0GradInstance = ck::tensor_operation::host::ReferenceBatchedGemm<InputDataType,
DataType, InputDataType,
DataType, InputDataType,
AccDataType, AccDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>; Scale>;
using ReferenceGemm1GradInstance = ck::tensor_operation::host::ReferenceBatchedGemm<InputDataType,
InputDataType,
OutputDataType,
AccDataType,
PassThrough,
PassThrough,
Scale>;
// Ref dropout // Ref dropout
using ReferenceDropoutInstance = using ReferenceDropoutInstance =
ck::tensor_operation::host::ReferenceDropout<ushort, DataType, DataType>; ck::tensor_operation::host::ReferenceDropout<ZDataType, InputDataType, InputDataType>;
template <typename TensorQ, template <typename TensorQ,
typename TensorK, typename TensorK,
...@@ -411,7 +434,7 @@ void run_attention_fwd_host(const TensorQ& q_g_m_k, ...@@ -411,7 +434,7 @@ void run_attention_fwd_host(const TensorQ& q_g_m_k,
TensorLSE& lse_g_m, TensorLSE& lse_g_m,
TensorP& p_drop_g_m_n, TensorP& p_drop_g_m_n,
TensorZ& z_g_m_n, TensorZ& z_g_m_n,
ushort p_dropout_in_16bits, ZDataType p_dropout_in_16bits,
float rp_dropout) float rp_dropout)
{ {
// S = alpha * Q * K^T // S = alpha * Q * K^T
...@@ -468,8 +491,8 @@ int run(int argc, char* argv[]) ...@@ -468,8 +491,8 @@ int run(int argc, char* argv[])
ck::index_t N = 512; ck::index_t N = 512;
ck::index_t K = DIM; ck::index_t K = DIM;
ck::index_t O = DIM; ck::index_t O = DIM;
ck::index_t G0 = 54; ck::index_t G0 = 4;
ck::index_t G1 = 16; ck::index_t G1 = 6;
bool input_permute = false; bool input_permute = false;
bool output_permute = false; bool output_permute = false;
...@@ -517,10 +540,10 @@ int run(int argc, char* argv[]) ...@@ -517,10 +540,10 @@ int run(int argc, char* argv[])
exit(0); exit(0);
} }
float p_dropout = 1 - p_drop; float p_dropout = 1 - p_drop;
uint16_t p_dropout_in_16bits = uint16_t(std::floor(p_dropout * 65535.0)); ZDataType p_dropout_in_16bits = ZDataType(std::floor(p_dropout * 65535.0));
float rp_dropout = 1.0 / p_dropout; float rp_dropout = 1.0 / p_dropout;
float alpha = 1.f / std::sqrt(K); float alpha = 1.f / std::sqrt(K);
std::cout << "do_verification: " << do_verification << std::endl; std::cout << "do_verification: " << do_verification << std::endl;
std::cout << "init_method: " << init_method << std::endl; std::cout << "init_method: " << init_method << std::endl;
...@@ -578,12 +601,12 @@ int run(int argc, char* argv[]) ...@@ -578,12 +601,12 @@ int run(int argc, char* argv[])
std::vector<ck::index_t> lse_gs_ms_lengths{G0, G1, M}; std::vector<ck::index_t> lse_gs_ms_lengths{G0, G1, M};
std::vector<ck::index_t> lse_gs_ms_strides{G1 * M, M, 1}; // LSE layout [G0, G1, M] std::vector<ck::index_t> lse_gs_ms_strides{G1 * M, M, 1}; // LSE layout [G0, G1, M]
Tensor<DataType> q_gs_ms_ks(q_gs_ms_ks_lengths, q_gs_ms_ks_strides); Tensor<InputDataType> q_gs_ms_ks(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
Tensor<DataType> k_gs_ns_ks(k_gs_ns_ks_lengths, k_gs_ns_ks_strides); Tensor<InputDataType> k_gs_ns_ks(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
Tensor<ZDataType> z_gs_ms_ns(z_gs_ms_ns_lengths, z_gs_ms_ns_strides); Tensor<ZDataType> z_gs_ms_ns(z_gs_ms_ns_lengths, z_gs_ms_ns_strides);
Tensor<DataType> v_gs_os_ns(v_gs_os_ns_lengths, v_gs_os_ns_strides); Tensor<InputDataType> v_gs_os_ns(v_gs_os_ns_lengths, v_gs_os_ns_strides);
Tensor<DataType> y_gs_ms_os(y_gs_ms_os_lengths, y_gs_ms_os_strides); Tensor<InputDataType> y_gs_ms_os(y_gs_ms_os_lengths, y_gs_ms_os_strides);
Tensor<DataType> ygrad_gs_ms_os(y_gs_ms_os_lengths, y_gs_ms_os_strides); Tensor<InputDataType> ygrad_gs_ms_os(y_gs_ms_os_lengths, y_gs_ms_os_strides);
Tensor<LSEDataType> lse_gs_ms(lse_gs_ms_lengths, lse_gs_ms_strides); Tensor<LSEDataType> lse_gs_ms(lse_gs_ms_lengths, lse_gs_ms_strides);
std::cout << "q_gs_ms_ks: " << q_gs_ms_ks.mDesc << std::endl; std::cout << "q_gs_ms_ks: " << q_gs_ms_ks.mDesc << std::endl;
...@@ -593,45 +616,45 @@ int run(int argc, char* argv[]) ...@@ -593,45 +616,45 @@ int run(int argc, char* argv[])
std::cout << "y_gs_ms_os: " << y_gs_ms_os.mDesc << std::endl; std::cout << "y_gs_ms_os: " << y_gs_ms_os.mDesc << std::endl;
std::cout << "lse_gs_ms_os: " << lse_gs_ms.mDesc << std::endl; std::cout << "lse_gs_ms_os: " << lse_gs_ms.mDesc << std::endl;
z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<DataType>{0}); z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<InputDataType>{0});
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
case 1: case 1:
q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2}); q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-2, 2});
k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2}); k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-2, 2});
v_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2}); v_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-2, 2});
ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2}); ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-2, 2});
break; break;
case 2: case 2:
q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<DataType>{0.0, 1.0}); q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<InputDataType>{0.0, 1.0});
k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<DataType>{0.0, 1.0}); k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<InputDataType>{0.0, 1.0});
v_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5}); v_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<InputDataType>{-0.5, 0.5});
ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5}); ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_3<InputDataType>{-0.5, 0.5});
break; break;
case 3: case 3:
q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5}); q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-5, 5});
k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
break; break;
case 4: case 4:
q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1});
k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1});
v_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); v_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1});
ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1});
break; break;
case 5: case 5:
q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1});
k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Sequential<2>{}); // dy[g0, g1, m, o] ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Sequential<2>{}); // dy[g0, g1, m, o]
// dO dot O = [0; 1; 2; ...] // dO dot O = [0; 1; 2; ...]
break; break;
case 6: case 6:
q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1});
k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Sequential<3>{}); // dy[g0, g1, m, o] ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Sequential<3>{}); // dy[g0, g1, m, o]
// assume mnko = 256 // assume mnko = 256
// P = softmax(QK) = 0.0039 * ones // P = softmax(QK) = 0.0039 * ones
...@@ -642,10 +665,10 @@ int run(int argc, char* argv[]) ...@@ -642,10 +665,10 @@ int run(int argc, char* argv[])
// //
break; break;
default: default:
q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1});
k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{}); v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<InputDataType>{});
ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); // dy[g0, g1, m, o] ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_1<InputDataType>{1}); // dy[g0, g1, m, o]
// assume mnko = 256 // assume mnko = 256
// P = softmax(QK) = 0.0039 * ones // P = softmax(QK) = 0.0039 * ones
// O = P V = 0.0039 * ones // O = P V = 0.0039 * ones
...@@ -656,14 +679,14 @@ int run(int argc, char* argv[]) ...@@ -656,14 +679,14 @@ int run(int argc, char* argv[])
// = 0 // = 0
} }
Tensor<DataType> q_g_m_k({BatchCount, M, K}); Tensor<InputDataType> q_g_m_k({BatchCount, M, K});
Tensor<DataType> k_g_n_k({BatchCount, N, K}); Tensor<InputDataType> k_g_n_k({BatchCount, N, K});
Tensor<ZDataType> z_g_m_n({BatchCount, M, N}); Tensor<ZDataType> z_g_m_n({BatchCount, M, N});
Tensor<DataType> v_g_n_o({BatchCount, N, O}); Tensor<InputDataType> v_g_n_o({BatchCount, N, O});
Tensor<AccDataType> s_g_m_n({BatchCount, M, N}); Tensor<AccDataType> s_g_m_n({BatchCount, M, N});
Tensor<DataType> p_g_m_n({BatchCount, M, N}); Tensor<InputDataType> p_g_m_n({BatchCount, M, N});
Tensor<DataType> p_drop_g_m_n({BatchCount, M, N}); Tensor<InputDataType> p_drop_g_m_n({BatchCount, M, N});
Tensor<DataType> y_g_m_o({BatchCount, M, O}); Tensor<InputDataType> y_g_m_o({BatchCount, M, O});
Tensor<LSEDataType> lse_g_m({BatchCount, M}); Tensor<LSEDataType> lse_g_m({BatchCount, M});
q_gs_ms_ks.ForEach( q_gs_ms_ks.ForEach(
...@@ -674,16 +697,16 @@ int run(int argc, char* argv[]) ...@@ -674,16 +697,16 @@ int run(int argc, char* argv[])
[&](auto& self, auto idx) { v_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx); }); [&](auto& self, auto idx) { v_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx); });
// qkv gradients have the same descriptor as with qkv // qkv gradients have the same descriptor as with qkv
DeviceMem q_device_buf(sizeof(DataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize()); DeviceMem q_device_buf(sizeof(InputDataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem k_device_buf(sizeof(DataType) * k_gs_ns_ks.mDesc.GetElementSpaceSize()); DeviceMem k_device_buf(sizeof(InputDataType) * k_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem z_device_buf(sizeof(ZDataType) * z_gs_ms_ns.mDesc.GetElementSpaceSize()); DeviceMem z_device_buf(sizeof(ZDataType) * z_gs_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem v_device_buf(sizeof(DataType) * v_gs_os_ns.mDesc.GetElementSpaceSize()); DeviceMem v_device_buf(sizeof(InputDataType) * v_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem y_device_buf(sizeof(DataType) * y_gs_ms_os.mDesc.GetElementSpaceSize()); DeviceMem y_device_buf(sizeof(InputDataType) * y_gs_ms_os.mDesc.GetElementSpaceSize());
DeviceMem lse_device_buf(sizeof(LSEDataType) * lse_gs_ms.mDesc.GetElementSpaceSize()); DeviceMem lse_device_buf(sizeof(LSEDataType) * lse_gs_ms.mDesc.GetElementSpaceSize());
DeviceMem qgrad_device_buf(sizeof(DataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize()); DeviceMem qgrad_device_buf(sizeof(OutputDataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem kgrad_device_buf(sizeof(DataType) * k_gs_ns_ks.mDesc.GetElementSpaceSize()); DeviceMem kgrad_device_buf(sizeof(OutputDataType) * k_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem vgrad_device_buf(sizeof(DataType) * v_gs_os_ns.mDesc.GetElementSpaceSize()); DeviceMem vgrad_device_buf(sizeof(OutputDataType) * v_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem ygrad_device_buf(sizeof(DataType) * y_gs_ms_os.mDesc.GetElementSpaceSize()); DeviceMem ygrad_device_buf(sizeof(InputDataType) * y_gs_ms_os.mDesc.GetElementSpaceSize());
q_device_buf.ToDevice(q_gs_ms_ks.mData.data()); q_device_buf.ToDevice(q_gs_ms_ks.mData.data());
k_device_buf.ToDevice(k_gs_ns_ks.mData.data()); k_device_buf.ToDevice(k_gs_ns_ks.mData.data());
...@@ -696,16 +719,16 @@ int run(int argc, char* argv[]) ...@@ -696,16 +719,16 @@ int run(int argc, char* argv[])
// get z matrix // get z matrix
{ {
auto argument = gemm.MakeArgument( auto argument = gemm.MakeArgument(
static_cast<DataType*>(q_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(q_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(k_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(k_device_buf.GetDeviceBuffer()),
static_cast<ZDataType*>(z_device_buf.GetDeviceBuffer()), static_cast<ZDataType*>(z_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(v_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(v_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(y_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(y_device_buf.GetDeviceBuffer()),
static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()), static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(ygrad_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(ygrad_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(qgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(qgrad_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(kgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(kgrad_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(vgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(vgrad_device_buf.GetDeviceBuffer()),
{}, // std::array<void*, 1> p_acc0_biases; {}, // std::array<void*, 1> p_acc0_biases;
{}, // std::array<void*, 1> p_acc1_biases; {}, // std::array<void*, 1> p_acc1_biases;
q_gs_ms_ks_lengths, q_gs_ms_ks_lengths,
...@@ -741,16 +764,16 @@ int run(int argc, char* argv[]) ...@@ -741,16 +764,16 @@ int run(int argc, char* argv[])
} }
// not need output z matrix // not need output z matrix
auto argument = gemm.MakeArgument( auto argument = gemm.MakeArgument(
static_cast<DataType*>(q_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(q_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(k_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(k_device_buf.GetDeviceBuffer()),
static_cast<ZDataType*>(nullptr), // set to nullptr static_cast<ZDataType*>(nullptr), // set to nullptr
static_cast<DataType*>(v_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(v_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(y_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(y_device_buf.GetDeviceBuffer()),
static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()), static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(ygrad_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(ygrad_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(qgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(qgrad_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(kgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(kgrad_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(vgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(vgrad_device_buf.GetDeviceBuffer()),
{}, // std::array<void*, 1> p_acc0_biases; {}, // std::array<void*, 1> p_acc0_biases;
{}, // std::array<void*, 1> p_acc1_biases; {}, // std::array<void*, 1> p_acc1_biases;
q_gs_ms_ks_lengths, q_gs_ms_ks_lengths,
...@@ -785,10 +808,13 @@ int run(int argc, char* argv[]) ...@@ -785,10 +808,13 @@ int run(int argc, char* argv[])
// 3x MNK + 2x MNO // 3x MNK + 2x MNO
std::size_t flop = (size_t(3) * M * N * K + size_t(2) * M * N * O) * 2 * BatchCount; std::size_t flop = (size_t(3) * M * N * K + size_t(2) * M * N * O) * 2 * BatchCount;
// Q/K/V/Y, dQ/dK/dV/dY, LSE // Q/K/V/Y, dQ/dK/dV/dY, LSE
std::size_t num_btype = (sizeof(DataType) * M * K + sizeof(DataType) * K * N + std::size_t num_btype =
sizeof(DataType) * N * O + sizeof(DataType) * M * O) * (sizeof(InputDataType) * M * K + sizeof(InputDataType) * K * N +
size_t(2) * BatchCount + sizeof(InputDataType) * N * O + sizeof(InputDataType) * M * O * size_t(2) +
sizeof(LSEDataType) * M * BatchCount; sizeof(OutputDataType) * M * K + sizeof(OutputDataType) * K * N +
sizeof(OutputDataType) * N * O) *
BatchCount +
sizeof(LSEDataType) * M * BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
...@@ -831,14 +857,14 @@ int run(int argc, char* argv[]) ...@@ -831,14 +857,14 @@ int run(int argc, char* argv[])
qgrad_device_buf.SetZero(); qgrad_device_buf.SetZero();
invoker.Run(argument, StreamConfig{nullptr, false}); invoker.Run(argument, StreamConfig{nullptr, false});
Tensor<DataType> qgrad_g_m_k({BatchCount, M, K}); Tensor<OutputDataType> qgrad_g_m_k({BatchCount, M, K});
Tensor<DataType> kgrad_g_n_k({BatchCount, N, K}); Tensor<OutputDataType> kgrad_g_n_k({BatchCount, N, K});
Tensor<DataType> vgrad_g_n_o({BatchCount, N, O}); Tensor<OutputDataType> vgrad_g_n_o({BatchCount, N, O});
Tensor<DataType> sgrad_g_m_n({BatchCount, M, N}); Tensor<InputDataType> sgrad_g_m_n({BatchCount, M, N});
Tensor<DataType> pgrad_g_m_n({BatchCount, M, N}); Tensor<InputDataType> pgrad_g_m_n({BatchCount, M, N});
Tensor<DataType> pgrad_drop_g_m_n({BatchCount, M, N}); Tensor<InputDataType> pgrad_drop_g_m_n({BatchCount, M, N});
Tensor<DataType> ygrad_g_m_o({BatchCount, M, O}); Tensor<InputDataType> ygrad_g_m_o({BatchCount, M, O});
Tensor<DataType> ygrad_dot_y_g_m({BatchCount, M}); Tensor<InputDataType> ygrad_dot_y_g_m({BatchCount, M});
ygrad_gs_ms_os.ForEach([&](auto& self, auto idx) { ygrad_gs_ms_os.ForEach([&](auto& self, auto idx) {
ygrad_g_m_o(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx); ygrad_g_m_o(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
...@@ -854,13 +880,16 @@ int run(int argc, char* argv[]) ...@@ -854,13 +880,16 @@ int run(int argc, char* argv[])
#endif #endif
// Gradients // Gradients
auto ref_gemm_grad = ReferenceGemmGradInstance{}; auto ref_gemm0_grad = ReferenceGemm0GradInstance{};
auto ref_gemm_grad_invoker = ref_gemm_grad.MakeInvoker(); auto ref_gemm0_grad_invoker = ref_gemm0_grad.MakeInvoker();
using RefGemmGradArg = ReferenceGemmGradInstance::Argument; using RefGemm0GradArg = ReferenceGemm0GradInstance::Argument;
auto ref_gemm1_grad = ReferenceGemm1GradInstance{};
auto ref_gemm1_grad_invoker = ref_gemm1_grad.MakeInvoker();
using RefGemm1GradArg = ReferenceGemm1GradInstance::Argument;
// dP_dropout = dY * V^T // dP_dropout = dY * V^T
auto v_g_o_n = v_g_n_o.Transpose({0, 2, 1}); auto v_g_o_n = v_g_n_o.Transpose({0, 2, 1});
ref_gemm_grad_invoker.Run(RefGemmGradArg{ ref_gemm0_grad_invoker.Run(RefGemm0GradArg{
ygrad_g_m_o, v_g_o_n, pgrad_drop_g_m_n, PassThrough{}, PassThrough{}, Scale{1.f}}); ygrad_g_m_o, v_g_o_n, pgrad_drop_g_m_n, PassThrough{}, PassThrough{}, Scale{1.f}});
#if PRINT_HOST #if PRINT_HOST
{ {
...@@ -887,7 +916,7 @@ int run(int argc, char* argv[]) ...@@ -887,7 +916,7 @@ int run(int argc, char* argv[])
ygrad_dot_y += ck::type_convert<AccDataType>(ygrad_g_m_o(idx_gmo)) * ygrad_dot_y += ck::type_convert<AccDataType>(ygrad_g_m_o(idx_gmo)) *
ck::type_convert<AccDataType>(y_g_m_o(idx_gmo)); ck::type_convert<AccDataType>(y_g_m_o(idx_gmo));
} }
self(idx_gmn) = ck::type_convert<DataType>( self(idx_gmn) = ck::type_convert<InputDataType>(
ck::type_convert<AccDataType>(p_g_m_n(idx_gmn)) * ck::type_convert<AccDataType>(p_g_m_n(idx_gmn)) *
(ck::type_convert<AccDataType>(pgrad_g_m_n(idx_gmn)) - ygrad_dot_y)); (ck::type_convert<AccDataType>(pgrad_g_m_n(idx_gmn)) - ygrad_dot_y));
}); });
...@@ -903,7 +932,7 @@ int run(int argc, char* argv[]) ...@@ -903,7 +932,7 @@ int run(int argc, char* argv[])
#endif #endif
// dV = P_drop^T * dY // dV = P_drop^T * dY
auto p_drop_g_n_m = p_drop_g_m_n.Transpose({0, 2, 1}); auto p_drop_g_n_m = p_drop_g_m_n.Transpose({0, 2, 1});
ref_gemm_grad_invoker.Run(RefGemmGradArg{ ref_gemm1_grad_invoker.Run(RefGemm1GradArg{
p_drop_g_n_m, ygrad_g_m_o, vgrad_g_n_o, PassThrough{}, PassThrough{}, Scale{1.0f}}); p_drop_g_n_m, ygrad_g_m_o, vgrad_g_n_o, PassThrough{}, PassThrough{}, Scale{1.0f}});
#if PRINT_HOST #if PRINT_HOST
{ {
...@@ -915,7 +944,7 @@ int run(int argc, char* argv[]) ...@@ -915,7 +944,7 @@ int run(int argc, char* argv[])
#endif #endif
// dQ = alpha * dS * K // dQ = alpha * dS * K
ref_gemm_grad_invoker.Run(RefGemmGradArg{ ref_gemm1_grad_invoker.Run(RefGemm1GradArg{
sgrad_g_m_n, k_g_n_k, qgrad_g_m_k, PassThrough{}, PassThrough{}, Scale{alpha}}); sgrad_g_m_n, k_g_n_k, qgrad_g_m_k, PassThrough{}, PassThrough{}, Scale{alpha}});
#if PRINT_HOST #if PRINT_HOST
{ {
...@@ -928,7 +957,7 @@ int run(int argc, char* argv[]) ...@@ -928,7 +957,7 @@ int run(int argc, char* argv[])
// dK = alpha * dS^T * Q // dK = alpha * dS^T * Q
auto sgrad_g_n_m = sgrad_g_m_n.Transpose({0, 2, 1}); auto sgrad_g_n_m = sgrad_g_m_n.Transpose({0, 2, 1});
ref_gemm_grad_invoker.Run(RefGemmGradArg{ ref_gemm1_grad_invoker.Run(RefGemm1GradArg{
sgrad_g_n_m, q_g_m_k, kgrad_g_n_k, PassThrough{}, PassThrough{}, Scale{alpha}}); sgrad_g_n_m, q_g_m_k, kgrad_g_n_k, PassThrough{}, PassThrough{}, Scale{alpha}});
#if PRINT_HOST #if PRINT_HOST
{ {
...@@ -939,13 +968,13 @@ int run(int argc, char* argv[]) ...@@ -939,13 +968,13 @@ int run(int argc, char* argv[])
} }
#endif #endif
Tensor<DataType> qgrad_gs_ms_ks_host_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides); Tensor<OutputDataType> qgrad_gs_ms_ks_host_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
Tensor<DataType> kgrad_gs_ns_ks_host_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides); Tensor<OutputDataType> kgrad_gs_ns_ks_host_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
Tensor<DataType> vgrad_gs_os_ns_host_result(v_gs_os_ns_lengths, v_gs_os_ns_strides); Tensor<OutputDataType> vgrad_gs_os_ns_host_result(v_gs_os_ns_lengths, v_gs_os_ns_strides);
Tensor<DataType> qgrad_gs_ms_ks_device_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides); Tensor<OutputDataType> qgrad_gs_ms_ks_device_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
Tensor<DataType> kgrad_gs_ns_ks_device_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides); Tensor<OutputDataType> kgrad_gs_ns_ks_device_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
Tensor<DataType> vgrad_gs_os_ns_device_result(v_gs_os_ns_lengths, v_gs_os_ns_strides); Tensor<OutputDataType> vgrad_gs_os_ns_device_result(v_gs_os_ns_lengths, v_gs_os_ns_strides);
qgrad_device_buf.FromDevice(qgrad_gs_ms_ks_device_result.mData.data()); qgrad_device_buf.FromDevice(qgrad_gs_ms_ks_device_result.mData.data());
kgrad_device_buf.FromDevice(kgrad_gs_ns_ks_device_result.mData.data()); kgrad_device_buf.FromDevice(kgrad_gs_ns_ks_device_result.mData.data());
......
include/ck/tensor_operation/gpu/device/impl/device_batched_multihead_attention_backward_xdl_cshuffle_v4.hpp
View file @ 1a306e0d
...@@ -28,7 +28,8 @@ namespace tensor_operation { ...@@ -28,7 +28,8 @@ namespace tensor_operation {
namespace device { namespace device {
template <typename GridwiseGemm, template <typename GridwiseGemm,
typename DataType, typename InputDataType,
typename OutputDataType,
typename ZDataType, typename ZDataType,
typename LSEDataType, typename LSEDataType,
typename AElementwiseOperation, typename AElementwiseOperation,
...@@ -46,22 +47,23 @@ template <typename GridwiseGemm, ...@@ -46,22 +47,23 @@ template <typename GridwiseGemm,
typename Block2CTileMap, typename Block2CTileMap,
typename ComputeBasePtrOfStridedBatch, typename ComputeBasePtrOfStridedBatch,
typename C0MatrixMask, typename C0MatrixMask,
bool HasMainKBlockLoop> bool HasMainKBlockLoop,
bool Deterministic>
__global__ void __global__ void
#if CK_USE_LAUNCH_BOUNDS #if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, /*CK_MIN_BLOCK_PER_CU*/ 1) __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, /*CK_MIN_BLOCK_PER_CU*/ 1)
#endif #endif
kernel_batched_multihead_attention_backward_xdl_cshuffle_v1( kernel_batched_multihead_attention_backward_xdl_cshuffle_v1(
const DataType* __restrict__ p_a_grid, const InputDataType* __restrict__ p_a_grid,
const DataType* __restrict__ p_b_grid, const InputDataType* __restrict__ p_b_grid,
ZDataType* __restrict__ p_z_grid, ZDataType* __restrict__ p_z_grid,
const DataType* __restrict__ p_b1_grid, const InputDataType* __restrict__ p_b1_grid,
const DataType* __restrict__ p_c_grid, const InputDataType* __restrict__ p_c_grid,
const LSEDataType* __restrict__ p_lse_grid, const LSEDataType* __restrict__ p_lse_grid,
const DataType* __restrict__ p_ygrad_grid, const InputDataType* __restrict__ p_ygrad_grid,
DataType* __restrict__ p_qgrad_grid, OutputDataType* __restrict__ p_qgrad_grid,
DataType* __restrict__ p_kgrad_grid, OutputDataType* __restrict__ p_kgrad_grid,
DataType* __restrict__ p_vgrad_grid, OutputDataType* __restrict__ p_vgrad_grid,
const AElementwiseOperation a_element_op, const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op, const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op, const AccElementwiseOperation acc_element_op,
...@@ -78,6 +80,7 @@ __global__ void ...@@ -78,6 +80,7 @@ __global__ void
const YGradGridDesc_O0_M_O1 ygrad_grid_desc_o0_m_o1, const YGradGridDesc_O0_M_O1 ygrad_grid_desc_o0_m_o1,
const Block2CTileMap block_2_ctile_map, const Block2CTileMap block_2_ctile_map,
const index_t batch_count, const index_t batch_count,
const index_t nblock,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch, const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask, const C0MatrixMask c0_matrix_mask,
const float p_drop, const float p_drop,
...@@ -109,33 +112,72 @@ __global__ void ...@@ -109,33 +112,72 @@ __global__ void
ck::philox ph(seed, global_thread_id, offset); ck::philox ph(seed, global_thread_id, offset);
ZDataType* z_matrix_ptr = (p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset); ZDataType* z_matrix_ptr = (p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset);
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset, if constexpr(Deterministic)
p_b_grid + b_batch_offset, {
z_matrix_ptr, for(index_t i = 0; i < nblock; i++)
p_b1_grid + b1_batch_offset, {
p_c_grid + c_batch_offset, GridwiseGemm::template Run<HasMainKBlockLoop>(
p_lse_grid + lse_batch_offset, p_a_grid + a_batch_offset,
p_ygrad_grid + c_batch_offset, p_b_grid + b_batch_offset,
p_qgrad_grid + a_batch_offset, z_matrix_ptr,
p_kgrad_grid + b_batch_offset, p_b1_grid + b1_batch_offset,
p_vgrad_grid + b1_batch_offset, p_c_grid + c_batch_offset,
p_shared, p_lse_grid + lse_batch_offset,
a_element_op, p_ygrad_grid + c_batch_offset,
b_element_op, p_qgrad_grid + a_batch_offset,
acc_element_op, p_kgrad_grid + b_batch_offset,
b1_element_op, p_vgrad_grid + b1_batch_offset,
c_element_op, p_shared,
a_grid_desc_ak0_m_ak1, a_element_op,
b_grid_desc_bk0_n_bk1, b_element_op,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, acc_element_op,
b1_grid_desc_bk0_n_bk1, b1_element_op,
c_grid_desc_mblock_mperblock_nblock_nperblock, c_element_op,
lse_grid_desc_m, a_grid_desc_ak0_m_ak1,
ygrad_grid_desc_o0_m_o1, b_grid_desc_bk0_n_bk1,
block_2_ctile_map, c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
c0_matrix_mask, b1_grid_desc_bk0_n_bk1,
p_drop, c_grid_desc_mblock_mperblock_nblock_nperblock,
ph); lse_grid_desc_m,
ygrad_grid_desc_o0_m_o1,
block_2_ctile_map,
c0_matrix_mask,
p_drop,
ph,
i);
}
}
else
{
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
z_matrix_ptr,
p_b1_grid + b1_batch_offset,
p_c_grid + c_batch_offset,
p_lse_grid + lse_batch_offset,
p_ygrad_grid + c_batch_offset,
p_qgrad_grid + a_batch_offset,
p_kgrad_grid + b_batch_offset,
p_vgrad_grid + b1_batch_offset,
p_shared,
a_element_op,
b_element_op,
acc_element_op,
b1_element_op,
c_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
b1_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
lse_grid_desc_m,
ygrad_grid_desc_o0_m_o1,
block_2_ctile_map,
c0_matrix_mask,
p_drop,
ph,
0);
}
#else #else
ignore = p_a_grid; ignore = p_a_grid;
ignore = p_b_grid; ignore = p_b_grid;
...@@ -168,7 +210,8 @@ template <index_t NumDimG, ...@@ -168,7 +210,8 @@ template <index_t NumDimG,
index_t NumDimN, index_t NumDimN,
index_t NumDimK, index_t NumDimK,
index_t NumDimO, // NumDimGemm1N index_t NumDimO, // NumDimGemm1N
typename DataType, typename InputDataType,
typename OutputDataType,
typename GemmDataType, typename GemmDataType,
typename ZDataType, typename ZDataType,
typename LSEDataType, typename LSEDataType,
...@@ -221,6 +264,7 @@ template <index_t NumDimG, ...@@ -221,6 +264,7 @@ template <index_t NumDimG,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
MaskingSpecialization MaskingSpec, MaskingSpecialization MaskingSpec,
bool Deterministic,
LoopScheduler LoopSched = LoopScheduler::Default> LoopScheduler LoopSched = LoopScheduler::Default>
struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
: public BaseOperator // TODO inherit atten bwd op once API stablizes : public BaseOperator // TODO inherit atten bwd op once API stablizes
...@@ -587,7 +631,9 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -587,7 +631,9 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
// GridwiseGemm // GridwiseGemm
using GridwiseGemm = GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1< using GridwiseGemm = GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1<
DataType, // TODO: distinguish A/B datatype InputDataType, // TODO: distinguish A/B datatype
OutputDataType,
ZDataType,
GemmDataType, GemmDataType,
GemmAccDataType, GemmAccDataType,
CShuffleDataType, CShuffleDataType,
...@@ -643,22 +689,23 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -643,22 +689,23 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
CShuffleBlockTransferScalarPerVector_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopSched, LoopSched,
Transform::matrix_padder.PadN, Transform::matrix_padder.PadN,
MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>; MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle,
Deterministic>;
// Argument // Argument
struct Argument : public BaseArgument struct Argument : public BaseArgument
{ {
Argument( Argument(
const DataType* p_a_grid, const InputDataType* p_a_grid,
const DataType* p_b_grid, const InputDataType* p_b_grid,
ZDataType* p_z_grid, ZDataType* p_z_grid,
const DataType* p_b1_grid, const InputDataType* p_b1_grid,
const DataType* p_c_grid, // for dS const InputDataType* p_c_grid, // for dS
const LSEDataType* p_lse_grid, const LSEDataType* p_lse_grid,
const DataType* p_ygrad_grid, const InputDataType* p_ygrad_grid,
DataType* p_qgrad_grid, OutputDataType* p_qgrad_grid,
DataType* p_kgrad_grid, OutputDataType* p_kgrad_grid,
DataType* p_vgrad_grid, OutputDataType* p_vgrad_grid,
const std::array<void*, NumAcc0Bias> p_acc0_biases, const std::array<void*, NumAcc0Bias> p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases, const std::array<void*, NumAcc1Bias> p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths, const std::vector<index_t>& a_gs_ms_ks_lengths,
...@@ -802,16 +849,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -802,16 +849,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
} }
// pointers // pointers
const DataType* p_a_grid_; const InputDataType* p_a_grid_;
const DataType* p_b_grid_; const InputDataType* p_b_grid_;
ZDataType* p_z_grid_; ZDataType* p_z_grid_;
const DataType* p_b1_grid_; const InputDataType* p_b1_grid_;
const DataType* p_c_grid_; const InputDataType* p_c_grid_;
const LSEDataType* p_lse_grid_; const LSEDataType* p_lse_grid_;
const DataType* p_ygrad_grid_; const InputDataType* p_ygrad_grid_;
DataType* p_qgrad_grid_; OutputDataType* p_qgrad_grid_;
DataType* p_kgrad_grid_; OutputDataType* p_kgrad_grid_;
DataType* p_vgrad_grid_; OutputDataType* p_vgrad_grid_;
// tensor descriptor // tensor descriptor
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_; AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
...@@ -876,14 +923,17 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -876,14 +923,17 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
} }
const index_t grid_size = const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.k_grid_desc_n_k_) * arg.batch_count_; (Deterministic ? 1
: arg.block_2_ctile_map_.CalculateGridSize(arg.k_grid_desc_n_k_)) *
arg.batch_count_;
float ave_time = 0; float ave_time = 0;
auto launch_kernel = [&](auto has_main_k_block_loop_) { auto launch_kernel = [&](auto has_main_k_block_loop_) {
const auto kernel = kernel_batched_multihead_attention_backward_xdl_cshuffle_v1< const auto kernel = kernel_batched_multihead_attention_backward_xdl_cshuffle_v1<
GridwiseGemm, GridwiseGemm,
DataType, InputDataType,
OutputDataType,
ZDataType, ZDataType,
LSEDataType, LSEDataType,
AElementwiseOperation, AElementwiseOperation,
...@@ -901,42 +951,45 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -901,42 +951,45 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
typename GridwiseGemm::DefaultBlock2CTileMap, typename GridwiseGemm::DefaultBlock2CTileMap,
ComputeBasePtrOfStridedBatch, ComputeBasePtrOfStridedBatch,
C0MatrixMask, C0MatrixMask,
has_main_k_block_loop_>; has_main_k_block_loop_,
Deterministic>;
return launch_and_time_kernel(stream_config,
kernel, return launch_and_time_kernel(
dim3(grid_size), stream_config,
dim3(BlockSize), kernel,
0, dim3(grid_size),
arg.p_a_grid_, dim3(BlockSize),
arg.p_b_grid_, 0,
arg.p_z_grid_, arg.p_a_grid_,
arg.p_b1_grid_, arg.p_b_grid_,
arg.p_c_grid_, arg.p_z_grid_,
arg.p_lse_grid_, arg.p_b1_grid_,
arg.p_ygrad_grid_, arg.p_c_grid_,
arg.p_qgrad_grid_, arg.p_lse_grid_,
arg.p_kgrad_grid_, arg.p_ygrad_grid_,
arg.p_vgrad_grid_, arg.p_qgrad_grid_,
arg.a_element_op_, arg.p_kgrad_grid_,
arg.b_element_op_, arg.p_vgrad_grid_,
arg.acc_element_op_, arg.a_element_op_,
arg.b1_element_op_, arg.b_element_op_,
arg.c_element_op_, arg.acc_element_op_,
arg.a_grid_desc_ak0_m_ak1_, arg.b1_element_op_,
arg.b_grid_desc_bk0_n_bk1_, arg.c_element_op_,
arg.c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_, arg.a_grid_desc_ak0_m_ak1_,
arg.b1_grid_desc_bk0_n_bk1_, arg.b_grid_desc_bk0_n_bk1_,
arg.y_grid_desc_mblock_mperblock_oblock_operblock_, arg.c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
arg.lse_grid_desc_m_, arg.b1_grid_desc_bk0_n_bk1_,
arg.ygrad_grid_desc_o0_m_o1_, arg.y_grid_desc_mblock_mperblock_oblock_operblock_,
arg.block_2_ctile_map_, arg.lse_grid_desc_m_,
arg.batch_count_, arg.ygrad_grid_desc_o0_m_o1_,
arg.compute_base_ptr_of_batch_, arg.block_2_ctile_map_,
arg.c0_matrix_mask_, arg.batch_count_,
arg.p_drop_, arg.block_2_ctile_map_.CalculateGridSize(arg.k_grid_desc_n_k_),
arg.seed_, arg.compute_base_ptr_of_batch_,
arg.offset_); arg.c0_matrix_mask_,
arg.p_drop_,
arg.seed_,
arg.offset_);
}; };
// Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
...@@ -1041,16 +1094,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1041,16 +1094,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
} }
static auto MakeArgument( static auto MakeArgument(
const DataType* p_a, const InputDataType* p_a,
const DataType* p_b, const InputDataType* p_b,
ZDataType* p_z, ZDataType* p_z,
const DataType* p_b1, const InputDataType* p_b1,
const DataType* p_c, const InputDataType* p_c,
const LSEDataType* p_lse, const LSEDataType* p_lse,
const DataType* p_ygrad_grid, const InputDataType* p_ygrad_grid,
DataType* p_qgrad_grid, OutputDataType* p_qgrad_grid,
DataType* p_kgrad_grid, OutputDataType* p_kgrad_grid,
DataType* p_vgrad_grid, OutputDataType* p_vgrad_grid,
const std::array<void*, NumAcc0Bias> p_acc0_biases, const std::array<void*, NumAcc0Bias> p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases, const std::array<void*, NumAcc1Bias> p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths, const std::vector<index_t>& a_gs_ms_ks_lengths,
...@@ -1156,16 +1209,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1156,16 +1209,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
float p_drop, float p_drop,
std::tuple<unsigned long long, unsigned long long> seeds) // override std::tuple<unsigned long long, unsigned long long> seeds) // override
{ {
return std::make_unique<Argument>(static_cast<const DataType*>(p_a), return std::make_unique<Argument>(static_cast<const InputDataType*>(p_a),
static_cast<const DataType*>(p_b), static_cast<const InputDataType*>(p_b),
static_cast<ZDataType*>(p_z), static_cast<ZDataType*>(p_z),
static_cast<const DataType*>(p_b1), static_cast<const InputDataType*>(p_b1),
static_cast<const DataType*>(p_c), static_cast<const InputDataType*>(p_c),
static_cast<const LSEDataType*>(p_lse), static_cast<const LSEDataType*>(p_lse),
static_cast<const DataType*>(p_ygrad_grid), static_cast<const InputDataType*>(p_ygrad_grid),
static_cast<DataType*>(p_qgrad_grid), static_cast<OutputDataType*>(p_qgrad_grid),
static_cast<DataType*>(p_kgrad_grid), static_cast<OutputDataType*>(p_kgrad_grid),
static_cast<DataType*>(p_vgrad_grid), static_cast<OutputDataType*>(p_vgrad_grid),
p_acc0_biases, // cast in struct Argument p_acc0_biases, // cast in struct Argument
p_acc1_biases, // cast in struct Argument p_acc1_biases, // cast in struct Argument
a_gs_ms_ks_lengths, a_gs_ms_ks_lengths,
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt4.hpp
View file @ 1a306e0d
...@@ -20,7 +20,9 @@ ...@@ -20,7 +20,9 @@
namespace ck { namespace ck {
template <typename DataType, template <typename InputDataType,
typename OutputDataType,
typename ZDataType,
typename GemmDataType, typename GemmDataType,
typename FloatGemmAcc, typename FloatGemmAcc,
typename FloatCShuffle, typename FloatCShuffle,
...@@ -77,6 +79,7 @@ template <typename DataType, ...@@ -77,6 +79,7 @@ template <typename DataType,
LoopScheduler LoopSched, LoopScheduler LoopSched,
bool PadN, bool PadN,
bool MaskOutUpperTriangle, bool MaskOutUpperTriangle,
bool Deterministic,
PipelineVersion PipelineVer = PipelineVersion::v1> PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
{ {
...@@ -424,7 +427,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -424,7 +427,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
Sequence<AK0, MPerBlock, AK1>, Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder, ABlockTransferThreadClusterArrangeOrder,
DataType, InputDataType,
GemmDataType, GemmDataType,
GridDesc_K0_M_K1, GridDesc_K0_M_K1,
decltype(q_block_desc_k0_m_k1), decltype(q_block_desc_k0_m_k1),
...@@ -449,7 +452,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -449,7 +452,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
Sequence<BK0, NPerBlock, BK1>, Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder, BBlockTransferThreadClusterArrangeOrder,
DataType, InputDataType,
GemmDataType, GemmDataType,
GridDesc_K0_N_K1, GridDesc_K0_N_K1,
decltype(k_block_desc_k0_n_k1), decltype(k_block_desc_k0_n_k1),
...@@ -474,7 +477,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -474,7 +477,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
Sequence<BK0, NPerBlock, BK1>, Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder, BBlockTransferThreadClusterArrangeOrder,
DataType, InputDataType,
GemmDataType, GemmDataType,
GridDesc_K0_N_K1, GridDesc_K0_N_K1,
decltype(v_block_desc_k0_n_k1), decltype(v_block_desc_k0_n_k1),
...@@ -499,7 +502,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -499,7 +502,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
Sequence<AK0, MPerBlock, AK1>, Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder, ABlockTransferThreadClusterArrangeOrder,
DataType, InputDataType,
GemmDataType, GemmDataType,
GridDesc_K0_M_K1, GridDesc_K0_M_K1,
decltype(ygrad_block_desc_k0_m_k1), decltype(ygrad_block_desc_k0_m_k1),
...@@ -1080,7 +1083,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1080,7 +1083,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
typename ElementwiseOp = tensor_operation::element_wise::PassThrough> typename ElementwiseOp = tensor_operation::element_wise::PassThrough>
using CBlockwiseCopy = ThreadwiseTensorSliceTransfer_v1r3< using CBlockwiseCopy = ThreadwiseTensorSliceTransfer_v1r3<
FloatGemmAcc, FloatGemmAcc,
DataType, OutputDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4), decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
CGridDesc_M0_N0_M1_N1_M2_N2_N3_N4, CGridDesc_M0_N0_M1_N1_M2_N2_N3_N4,
ElementwiseOp, // CElementwiseOperation ElementwiseOp, // CElementwiseOperation
...@@ -1096,7 +1099,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1096,7 +1099,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
template <index_t BlockSize_, index_t BlockSliceLength_M_, index_t BlockSliceLength_O_> template <index_t BlockSize_, index_t BlockSliceLength_M_, index_t BlockSliceLength_O_>
struct YDotYGrad_M_O_ struct YDotYGrad_M_O_
{ {
static constexpr index_t SrcScalarPerVector = 16 / sizeof(DataType); static constexpr index_t SrcScalarPerVector = 16 / sizeof(InputDataType);
static constexpr auto ThreadClusterLength_O = static constexpr auto ThreadClusterLength_O =
Number<BlockSliceLength_O_ / SrcScalarPerVector>{}; Number<BlockSliceLength_O_ / SrcScalarPerVector>{};
static constexpr auto ThreadClusterLength_M = Number<BlockSize_ / ThreadClusterLength_O>{}; static constexpr auto ThreadClusterLength_M = Number<BlockSize_ / ThreadClusterLength_O>{};
...@@ -1213,16 +1216,16 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1213,16 +1216,16 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
typename Block2CTileMap, typename Block2CTileMap,
typename C0MatrixMask, typename C0MatrixMask,
typename YGradGridDesc_O0_M_O1> typename YGradGridDesc_O0_M_O1>
__device__ static void Run(const DataType* __restrict__ p_q_grid, __device__ static void Run(const InputDataType* __restrict__ p_q_grid,
const DataType* __restrict__ p_k_grid, const InputDataType* __restrict__ p_k_grid,
unsigned short* __restrict__ p_z_grid, ZDataType* __restrict__ p_z_grid,
const DataType* __restrict__ p_v_grid, const InputDataType* __restrict__ p_v_grid,
const DataType* __restrict__ p_y_grid, const InputDataType* __restrict__ p_y_grid,
const FloatLSE* __restrict__ p_lse_grid, const FloatLSE* __restrict__ p_lse_grid,
const DataType* __restrict__ p_ygrad_grid, const InputDataType* __restrict__ p_ygrad_grid,
DataType* __restrict__ p_qgrad_grid, OutputDataType* __restrict__ p_qgrad_grid,
DataType* __restrict__ p_kgrad_grid, OutputDataType* __restrict__ p_kgrad_grid,
DataType* __restrict__ p_vgrad_grid, OutputDataType* __restrict__ p_vgrad_grid,
void* __restrict__ p_shared, void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op, const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op, const BElementwiseOperation& b_element_op,
...@@ -1241,7 +1244,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1241,7 +1244,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
const Block2CTileMap& block_2_ctile_map, const Block2CTileMap& block_2_ctile_map,
const C0MatrixMask& c0_matrix_mask, const C0MatrixMask& c0_matrix_mask,
const float p_drop, const float p_drop,
ck::philox& ph) ck::philox& ph,
const index_t block_idx_n)
{ {
const FloatGemmAcc p_dropout = type_convert<FloatGemmAcc>(1.0f - p_drop); const FloatGemmAcc p_dropout = type_convert<FloatGemmAcc>(1.0f - p_drop);
const FloatGemmAcc rp_dropout = type_convert<FloatGemmAcc>(1.0f / p_dropout); const FloatGemmAcc rp_dropout = type_convert<FloatGemmAcc>(1.0f / p_dropout);
...@@ -1273,9 +1277,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1273,9 +1277,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
const auto block_work_idx = const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id())); block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
const index_t block_work_idx_n = Deterministic ? block_idx_n : block_work_idx[I0];
// HACK: this force n_block_data_idx_on_grid into SGPR // HACK: this force n_block_data_idx_on_grid into SGPR
const index_t n_block_data_idx_on_grid = const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I0] * NPerBlock); __builtin_amdgcn_readfirstlane(block_work_idx_n * NPerBlock);
// 6 GEMM operations are categorized into 3 buckets. SizeK == SizeO == head_dim // 6 GEMM operations are categorized into 3 buckets. SizeK == SizeO == head_dim
// S_MNK / dP_MNO Gemm (Gemm0 rcr) // S_MNK / dP_MNO Gemm (Gemm0 rcr)
...@@ -1605,7 +1611,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1605,7 +1611,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3< auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
ushort, ushort,
ushort, ZDataType,
decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5), decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5), decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
tensor_operation::element_wise::PassThrough, tensor_operation::element_wise::PassThrough,
...@@ -1625,15 +1631,15 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1625,15 +1631,15 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
1, // DstScalarStrideInVector 1, // DstScalarStrideInVector
true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
make_multi_index(block_work_idx[I0], // MBlockId make_multi_index(block_work_idx_n, // MBlockId
0, // NBlockId 0, // NBlockId
0, // mrepeat 0, // mrepeat
0, // nrepeat 0, // nrepeat
wave_id[I0], // MWaveId wave_id[I0], // MWaveId
wave_id[I1], // NWaveId wave_id[I1], // NWaveId
wave_m_n_id[I1], // MPerXdl wave_m_n_id[I1], // MPerXdl
0, // group 0, // group
wave_m_n_id[I0], // NInputIndex wave_m_n_id[I0], // NInputIndex
0), 0),
tensor_operation::element_wise::PassThrough{}}; tensor_operation::element_wise::PassThrough{}};
...@@ -1678,7 +1684,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1678,7 +1684,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
// performs for y // performs for y
auto y_threadwise_copy = ThreadwiseTensorSliceTransfer_v2< auto y_threadwise_copy = ThreadwiseTensorSliceTransfer_v2<
DataType, InputDataType,
FloatGemmAcc, FloatGemmAcc,
YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock, YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock,
decltype(y_thread_desc_m0_m1_o0_o1), decltype(y_thread_desc_m0_m1_o0_o1),
...@@ -1743,6 +1749,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1743,6 +1749,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
const index_t num_gemm0_m_block_outer_loop = q_grid_desc_k0_m_k1.GetLength(I1) / MPerBlock; const index_t num_gemm0_m_block_outer_loop = q_grid_desc_k0_m_k1.GetLength(I1) / MPerBlock;
constexpr index_t num_gemm1_k_block_inner_loop = MPerBlock / Gemm1KPerBlock; constexpr index_t num_gemm1_k_block_inner_loop = MPerBlock / Gemm1KPerBlock;
if constexpr(Deterministic)
{
block_sync_lds();
}
// Initialize dK&dV // Initialize dK&dV
kgrad_thread_buf.Clear(); kgrad_thread_buf.Clear();
vgrad_thread_buf.Clear(); vgrad_thread_buf.Clear();
...@@ -2286,7 +2297,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -2286,7 +2297,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder, Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
FloatCShuffle, // typename SrcData, FloatCShuffle, // typename SrcData,
DataType, // typename DstData, OutputDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock), decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(vgrad_grid_desc_nblock_nperblock_oblock_operblock), decltype(vgrad_grid_desc_nblock_nperblock_oblock_operblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder, Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
...@@ -2297,7 +2308,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -2297,7 +2308,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock, {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0), make_multi_index(0, 0, 0, 0),
vgrad_grid_desc_nblock_nperblock_oblock_operblock, vgrad_grid_desc_nblock_nperblock_oblock_operblock,
make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0), make_multi_index(block_work_idx_n, 0, block_work_idx[I1], 0),
c_element_op}; c_element_op};
// shuffle: threadwise copy C from VGPR to LDS // shuffle: threadwise copy C from VGPR to LDS
...@@ -2344,7 +2355,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -2344,7 +2355,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder, Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
FloatCShuffle, // typename SrcData, FloatCShuffle, // typename SrcData,
DataType, // typename DstData, OutputDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock), decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(kgrad_grid_desc_nblock_nperblock_oblock_operblock), decltype(kgrad_grid_desc_nblock_nperblock_oblock_operblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder, Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
...@@ -2355,7 +2366,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -2355,7 +2366,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock, {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0), make_multi_index(0, 0, 0, 0),
kgrad_grid_desc_nblock_nperblock_oblock_operblock, kgrad_grid_desc_nblock_nperblock_oblock_operblock,
make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0), make_multi_index(block_work_idx_n, 0, block_work_idx[I1], 0),
c_element_op}; c_element_op};
// space filling curve for threadwise C in VGPR // space filling curve for threadwise C in VGPR
......
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