Skip to content

GitLab

Commit 104aeabc authored by danyao12's avatar danyao12
Browse files

split bwd mqa/gqa

parent 198558c5
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 482 additions and 114 deletions
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_v2.cpp
View file @ 104aeabc
......@@ -542,8 +542,8 @@ int run(int argc, char* argv[])
static_cast<OutputDataType*>(qgrad_device_buf.GetDeviceBuffer()),
static_cast<OutputDataType*>(kgrad_device_buf.GetDeviceBuffer()),
static_cast<OutputDataType*>(vgrad_device_buf.GetDeviceBuffer()),
nullptr, // p_acc0_bias;
nullptr, // p_acc1_bias;
nullptr, // p_acc0_bias;
nullptr, // p_acc1_bias;
nullptr,
nullptr,
q_gs_ms_ks_lengths,
......@@ -593,8 +593,8 @@ int run(int argc, char* argv[])
static_cast<OutputDataType*>(qgrad_device_buf.GetDeviceBuffer()),
static_cast<OutputDataType*>(kgrad_device_buf.GetDeviceBuffer()),
static_cast<OutputDataType*>(vgrad_device_buf.GetDeviceBuffer()),
nullptr, // p_acc0_bias;
nullptr, // p_acc1_bias;
nullptr, // p_acc0_bias;
nullptr, // p_acc1_bias;
nullptr,
nullptr,
q_gs_ms_ks_lengths,
......
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_v3.cpp
View file @ 104aeabc
......@@ -277,7 +277,8 @@ int run(int argc, char* argv[])
ck::index_t K = DIM;
ck::index_t O = DIM;
ck::index_t G0 = 4;
ck::index_t G1 = 6;
ck::index_t G1 = 6; // h_q
ck::index_t G2 = 6; // h_kv
bool input_permute = false;
bool output_permute = false;
......@@ -296,7 +297,7 @@ int run(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 13)
else if(argc == 14)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
......@@ -308,20 +309,21 @@ int run(int argc, char* argv[])
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
G2 = std::stoi(argv[10]);
p_drop = std::stof(argv[10]);
p_drop = std::stof(argv[11]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
input_permute = std::stoi(argv[12]);
output_permute = std::stoi(argv[13]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
printf("arg4 to 10: M, N, K, O, G0, G1, G2\n");
printf("arg11: p_drop\n");
printf("arg12 to 13: input / output permute\n");
exit(0);
}
......@@ -339,6 +341,7 @@ int run(int argc, char* argv[])
std::cout << "O: " << O << std::endl;
std::cout << "G0: " << G0 << std::endl;
std::cout << "G1: " << G1 << std::endl;
std::cout << "G2: " << G2 << std::endl;
std::cout << "alpha: " << alpha << std::endl;
std::cout << "input_permute: " << input_permute << std::endl;
std::cout << "output_permute: " << output_permute << std::endl;
......@@ -354,17 +357,17 @@ int run(int argc, char* argv[])
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // Q layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // Q layout [G0, G1, M, K]
std::vector<ck::index_t> k_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> k_gs_ns_ks_lengths{G0, G2, N, K};
std::vector<ck::index_t> k_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // K layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // K layout [G0, G1, N, K]
? std::vector<ck::index_t>{N * G2 * K, K, G2 * K, 1} // K layout [G0, N, G2, K]
: std::vector<ck::index_t>{G2 * N * K, N * K, K, 1}; // K layout [G0, G2, N, K]
std::vector<ck::index_t> v_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> v_gs_os_ns_lengths{G0, G2, O, N};
std::vector<ck::index_t> v_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // V layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // V layout [G0, G1, N, O]
? std::vector<ck::index_t>{N * G2 * O, O, 1, G2 * O} // V layout [G0, N, G2, O]
: std::vector<ck::index_t>{G2 * N * O, N * O, 1, O}; // V layout [G0, G2, N, O]
std::vector<ck::index_t> y_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> y_gs_ms_os_strides =
......@@ -377,6 +380,18 @@ int run(int argc, char* argv[])
input_permute
? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // Z layout [G0, M, G1, N]
: std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // Z layout [G0, G1, M, N]
std::vector<ck::index_t> kgrad_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> kgrad_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // KGrad layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // KGrad layout [G0, G1, N, K]
std::vector<ck::index_t> vgrad_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> vgrad_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // VGrad layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // VGrad layout [G0, G1, N, O]
// The softmax stat log-sum-exp (LSE) is used to speed up softmax calculation in backward pass
// Pi = exp(Si) / sum(exp(S0) + exp(S1) + ...)
// = exp(Si) / exp(log(sum(exp() + ...)))
......@@ -394,6 +409,8 @@ int run(int argc, char* argv[])
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<DDataType> d_gs_ms(lse_gs_ms_lengths, lse_gs_ms_strides);
Tensor<OutputDataType> kgrad_gs_ns_ks(kgrad_gs_ns_ks_lengths, kgrad_gs_ns_ks_strides);
Tensor<OutputDataType> vgrad_gs_os_ns(vgrad_gs_os_ns_lengths, vgrad_gs_os_ns_strides);
std::cout << "q_gs_ms_ks: " << q_gs_ms_ks.mDesc << std::endl;
std::cout << "k_gs_ns_ks: " << k_gs_ns_ks.mDesc << std::endl;
......@@ -402,6 +419,8 @@ int run(int argc, char* argv[])
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 << "d_gs_ms_os: " << d_gs_ms.mDesc << std::endl;
std::cout << "kgrad_gs_ns_ks: " << kgrad_gs_ns_ks.mDesc << std::endl;
std::cout << "vgrad_gs_os_ns: " << vgrad_gs_os_ns.mDesc << std::endl;
z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<InputDataType>{0});
switch(init_method)
......@@ -478,10 +497,20 @@ int run(int argc, char* argv[])
q_gs_ms_ks.ForEach(
[&](auto& self, auto idx) { q_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx); });
k_gs_ns_ks.ForEach(
[&](auto& self, auto idx) { k_g_n_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx); });
v_gs_os_ns.ForEach(
[&](auto& self, auto idx) { v_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx); });
k_g_n_k.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0] / G1;
const size_t& g1 = idx[0] % G1;
const size_t& g2 = g1 / (G1 / G2);
self(idx) = k_gs_ns_ks(g0, g2, idx[1], idx[2]);
});
v_g_n_o.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0] / G1;
const size_t& g1 = idx[0] % G1;
const size_t& g2 = g1 / (G1 / G2);
self(idx) = v_gs_os_ns(g0, g2, idx[2], idx[1]);
});
// qkv gradients have the same descriptor as with qkv
DeviceMem q_device_buf(sizeof(InputDataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize());
......@@ -491,8 +520,8 @@ int run(int argc, char* argv[])
DeviceMem y_device_buf(sizeof(InputDataType) * y_gs_ms_os.mDesc.GetElementSpaceSize());
DeviceMem lse_device_buf(sizeof(LSEDataType) * lse_gs_ms.mDesc.GetElementSpaceSize());
DeviceMem qgrad_device_buf(sizeof(OutputDataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem kgrad_device_buf(sizeof(OutputDataType) * k_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem vgrad_device_buf(sizeof(OutputDataType) * v_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem kgrad_device_buf(sizeof(OutputDataType) * kgrad_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem vgrad_device_buf(sizeof(OutputDataType) * vgrad_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem ygrad_device_buf(sizeof(InputDataType) * y_gs_ms_os.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms.mDesc.GetElementSpaceSize());
......@@ -533,6 +562,10 @@ int run(int argc, char* argv[])
y_gs_ms_os_lengths,
y_gs_ms_os_strides,
lse_gs_ms_lengths,
kgrad_gs_ns_ks_lengths,
kgrad_gs_ns_ks_strides,
vgrad_gs_os_ns_lengths,
vgrad_gs_os_ns_strides,
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
......@@ -581,6 +614,10 @@ int run(int argc, char* argv[])
y_gs_ms_os_lengths,
y_gs_ms_os_strides,
lse_gs_ms_lengths,
kgrad_gs_ns_ks_lengths,
kgrad_gs_ns_ks_strides,
vgrad_gs_os_ns_lengths,
vgrad_gs_os_ns_strides,
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
......@@ -763,12 +800,16 @@ int run(int argc, char* argv[])
#endif
Tensor<OutputDataType> qgrad_gs_ms_ks_host_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
Tensor<OutputDataType> kgrad_gs_ns_ks_host_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
Tensor<OutputDataType> vgrad_gs_os_ns_host_result(v_gs_os_ns_lengths, v_gs_os_ns_strides);
Tensor<OutputDataType> kgrad_gs_ns_ks_host_result(kgrad_gs_ns_ks_lengths,
kgrad_gs_ns_ks_strides);
Tensor<OutputDataType> vgrad_gs_os_ns_host_result(vgrad_gs_os_ns_lengths,
vgrad_gs_os_ns_strides);
Tensor<OutputDataType> qgrad_gs_ms_ks_device_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
Tensor<OutputDataType> kgrad_gs_ns_ks_device_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
Tensor<OutputDataType> vgrad_gs_os_ns_device_result(v_gs_os_ns_lengths, v_gs_os_ns_strides);
Tensor<OutputDataType> kgrad_gs_ns_ks_device_result(kgrad_gs_ns_ks_lengths,
kgrad_gs_ns_ks_strides);
Tensor<OutputDataType> vgrad_gs_os_ns_device_result(vgrad_gs_os_ns_lengths,
vgrad_gs_os_ns_strides);
qgrad_device_buf.FromDevice(qgrad_gs_ms_ks_device_result.mData.data());
kgrad_device_buf.FromDevice(kgrad_gs_ns_ks_device_result.mData.data());
......
example/32_batched_gemm_scale_softmax_gemm/grouped_multihead_attention_backward_v3.cpp
View file @ 104aeabc
......@@ -24,7 +24,7 @@ Kernel outputs:
*/
#define USING_MASK 0
#define DIM 32 // DIM should be a multiple of 8.
#define DIM 128 // DIM should be a multiple of 8.
#include <iostream>
#include <numeric>
......@@ -273,6 +273,7 @@ int run(int argc, char* argv[])
// y_g0_m_g1_o = permute(y_g0_g1_m_o, [0, 2, 1, 3])
float alpha = 1.f / std::sqrt(DIM);
float p_drop = 0.0;
int h_ratio = 1; // G1 / G2
bool input_permute = true;
bool output_permute = true;
......@@ -290,25 +291,26 @@ int run(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 7)
else if(argc == 8)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
p_drop = std::stof(argv[4]);
p_drop = std::stof(argv[4]);
h_ratio = std::stof(argv[5]);
input_permute = std::stoi(argv[5]);
output_permute = std::stoi(argv[6]);
input_permute = std::stoi(argv[6]);
output_permute = std::stoi(argv[7]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
printf("arg4: p_drop\n");
printf("arg5: h_ratio\n");
printf("arg6 to 7: input / output permute\n");
exit(0);
}
......@@ -376,24 +378,25 @@ int run(int argc, char* argv[])
int K = DIM;
int O = DIM;
int G0 = rand() % 4 + 1;
int G1 = rand() % 4 + 1;
int G2 = rand() % 4 + 1;
int G1 = G2 * h_ratio;
std::vector<ck::index_t> q_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> q_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // Q layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // Q layout [G0, G1, M, K]
std::vector<ck::index_t> k_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> k_gs_ns_ks_lengths{G0, G2, N, K};
std::vector<ck::index_t> k_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // K layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // K layout [G0, G1, N, K]
? std::vector<ck::index_t>{N * G2 * K, K, G2 * K, 1} // K layout [G0, N, G2, K]
: std::vector<ck::index_t>{G2 * N * K, N * K, K, 1}; // K layout [G0, G2, N, K]
std::vector<ck::index_t> v_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> v_gs_os_ns_lengths{G0, G2, O, N};
std::vector<ck::index_t> v_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // V layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // V layout [G0, G1, N, O]
? std::vector<ck::index_t>{N * G2 * O, O, 1, G2 * O} // V layout [G0, N, G2, O]
: std::vector<ck::index_t>{G2 * N * O, N * O, 1, O}; // V layout [G0, G2, N, O]
std::vector<ck::index_t> y_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> y_gs_ms_os_strides =
......@@ -406,6 +409,17 @@ int run(int argc, char* argv[])
input_permute
? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // Z layout [G0, M, G1, N]
: std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // Z layout [G0, G1, M, N]
std::vector<ck::index_t> kgrad_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> kgrad_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // KGrad layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // KGrad layout [G0, G1, N, K]
std::vector<ck::index_t> vgrad_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> vgrad_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // VGrad layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // VGrad layout [G0, G1, N, O]
// The softmax stat log-sum-exp (LSE) is used to speed up softmax calculation in backward
// pass Pi = exp(Si) / sum(exp(S0) + exp(S1) + ...)
// = exp(Si) / exp(log(sum(exp() + ...)))
......@@ -427,6 +441,10 @@ int run(int argc, char* argv[])
y_gs_ms_os_strides,
lse_gs_ms_lengths,
lse_gs_ms_strides,
kgrad_gs_ns_ks_lengths,
kgrad_gs_ns_ks_strides,
vgrad_gs_os_ns_lengths,
vgrad_gs_os_ns_strides,
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
......@@ -451,6 +469,8 @@ int run(int argc, char* argv[])
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<DDataType> d_gs_ms(lse_gs_ms_lengths, lse_gs_ms_strides);
Tensor<OutputDataType> kgrad_gs_ns_ks(kgrad_gs_ns_ks_lengths, kgrad_gs_ns_ks_strides);
Tensor<OutputDataType> vgrad_gs_os_ns(vgrad_gs_os_ns_lengths, vgrad_gs_os_ns_strides);
if(i < 4)
{
std::cout << "q_gs_ms_ks: " << q_gs_ms_ks.mDesc << std::endl;
......@@ -460,6 +480,8 @@ int run(int argc, char* argv[])
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 << "d_gs_ms_os: " << d_gs_ms.mDesc << std::endl;
std::cout << "kgrad_gs_ns_ks: " << kgrad_gs_ns_ks.mDesc << std::endl;
std::cout << "vgrad_gs_os_ns: " << vgrad_gs_os_ns.mDesc << std::endl;
}
z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<InputDataType>{0});
switch(init_method)
......@@ -538,11 +560,19 @@ int run(int argc, char* argv[])
q_gs_ms_ks.ForEach([&](auto& self, auto idx) {
q_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
k_gs_ns_ks.ForEach([&](auto& self, auto idx) {
k_g_n_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
k_g_n_k.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0] / G1;
const size_t& g1 = idx[0] % G1;
const size_t& g2 = g1 / h_ratio;
self(idx) = k_gs_ns_ks(g0, g2, idx[1], idx[2]);
});
v_gs_os_ns.ForEach([&](auto& self, auto idx) {
v_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
v_g_n_o.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0] / G1;
const size_t& g1 = idx[0] % G1;
const size_t& g2 = g1 / h_ratio;
self(idx) = v_gs_os_ns(g0, g2, idx[2], idx[1]);
});
q_g_m_ks.push_back(q_g_m_k);
......@@ -562,6 +592,8 @@ int run(int argc, char* argv[])
z_tensors.push_back(z_gs_ms_ns);
lse_tensors.push_back(lse_gs_ms);
ygrad_tensors.push_back(ygrad_gs_ms_os);
kgrad_tensors.push_back(kgrad_gs_ns_ks);
vgrad_tensors.push_back(vgrad_gs_os_ns);
q_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(InputDataType) * q_gs_ms_ks.GetElementSpaceSize()));
k_tensors_device.emplace_back(
......@@ -578,10 +610,10 @@ int run(int argc, char* argv[])
std::make_unique<DeviceMem>(sizeof(DDataType) * d_gs_ms.GetElementSpaceSize()));
qgrad_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(OutputDataType) * q_gs_ms_ks.GetElementSpaceSize()));
kgrad_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(OutputDataType) * k_gs_ns_ks.GetElementSpaceSize()));
vgrad_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(OutputDataType) * v_gs_os_ns.GetElementSpaceSize()));
kgrad_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(OutputDataType) * kgrad_gs_ns_ks.GetElementSpaceSize()));
vgrad_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(OutputDataType) * vgrad_gs_os_ns.GetElementSpaceSize()));
ygrad_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(InputDataType) * y_gs_ms_os.GetElementSpaceSize()));
q_tensors_device.back()->ToDevice(q_gs_ms_ks.data());
......@@ -625,6 +657,7 @@ int run(int argc, char* argv[])
QKVElementOp{},
YElementOp{},
p_drop,
h_ratio,
std::tuple<unsigned long long, unsigned long long>(seed, offset));
DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
......@@ -674,6 +707,7 @@ int run(int argc, char* argv[])
QKVElementOp{},
YElementOp{},
p_drop,
h_ratio,
std::tuple<unsigned long long, unsigned long long>(seed, offset));
DeviceMem problem_desc_workspace_verify(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace_verify.GetDeviceBuffer());
......@@ -687,7 +721,7 @@ int run(int argc, char* argv[])
for(std::size_t i = 0; i < group_count; i++)
{
int G1 = v_tensors[i].GetLengths()[1];
int G1 = q_tensors[i].GetLengths()[1];
// copy z matirx data form device
z_tensors_device[i]->FromDevice(z_tensors[i].mData.data());
z_tensors[i].ForEach([&](auto& self, auto idx) {
......@@ -724,8 +758,8 @@ int run(int argc, char* argv[])
for(std::size_t i = 0; i < group_count; i++)
{
int G0 = v_tensors[i].GetLengths()[0];
int G1 = v_tensors[i].GetLengths()[1];
int G0 = q_tensors[i].GetLengths()[0];
int G1 = q_tensors[i].GetLengths()[1];
int O = v_tensors[i].GetLengths()[2];
int N = v_tensors[i].GetLengths()[3];
int M = q_tensors[i].GetLengths()[2];
......@@ -783,17 +817,17 @@ int run(int argc, char* argv[])
Tensor<OutputDataType> qgrad_gs_ms_ks_host_result(q_tensors[i].GetLengths(),
q_tensors[i].GetStrides());
Tensor<OutputDataType> kgrad_gs_ns_ks_host_result(k_tensors[i].GetLengths(),
k_tensors[i].GetStrides());
Tensor<OutputDataType> vgrad_gs_os_ns_host_result(v_tensors[i].GetLengths(),
v_tensors[i].GetStrides());
Tensor<OutputDataType> kgrad_gs_ns_ks_host_result(kgrad_tensors[i].GetLengths(),
kgrad_tensors[i].GetStrides());
Tensor<OutputDataType> vgrad_gs_os_ns_host_result(vgrad_tensors[i].GetLengths(),
vgrad_tensors[i].GetStrides());
Tensor<OutputDataType> qgrad_gs_ms_ks_device_result(q_tensors[i].GetLengths(),
q_tensors[i].GetStrides());
Tensor<OutputDataType> kgrad_gs_ns_ks_device_result(k_tensors[i].GetLengths(),
k_tensors[i].GetStrides());
Tensor<OutputDataType> vgrad_gs_os_ns_device_result(v_tensors[i].GetLengths(),
v_tensors[i].GetStrides());
Tensor<OutputDataType> kgrad_gs_ns_ks_device_result(kgrad_tensors[i].GetLengths(),
kgrad_tensors[i].GetStrides());
Tensor<OutputDataType> vgrad_gs_os_ns_device_result(vgrad_tensors[i].GetLengths(),
vgrad_tensors[i].GetStrides());
qgrad_tensors_device[i]->FromDevice(qgrad_gs_ms_ks_device_result.data());
kgrad_tensors_device[i]->FromDevice(kgrad_gs_ns_ks_device_result.data());
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_light_v2.hpp
View file @ 104aeabc
......@@ -132,14 +132,17 @@ __global__ void
const CElementwiseOperation c_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const BGridDesc_BK0_N_BK1 bgrad_grid_desc_bk0_n_bk1,
const D0GridDescriptor_M0_N0_M1_M2_N1_M3 d0_grid_desc_m0_n0_m1_m2_n1_m3,
const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
const B1GridDesc_BK0_N_BK1 b1grad_grid_desc_bk0_n_bk1,
const LSEGridDescriptor_M lse_grid_desc_m,
const YGradGridDesc_M0_O_M1 ygrad_grid_desc_m0_o_m1,
const Block2CTileMap block_2_ctile_map,
const index_t batch_count,
const index_t h_ratio,
const index_t nblock,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask,
......@@ -155,21 +158,26 @@ __global__ void
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const index_t gkv_idx = __builtin_amdgcn_readfirstlane(g_idx / h_ratio);
// NOTE: assumes QKVY has the same layout as dQ/dK/dV/dY therefore being able to reuse batch
// offsets
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(gkv_idx)));
const long_index_t z_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetZBasePtr(g_idx)));
const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(gkv_idx)));
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetLSEBasePtr(g_idx)));
const long_index_t bgrad_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetBGradBasePtr(g_idx)));
const long_index_t b1grad_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1GradBasePtr(g_idx)));
ck::philox ph(seed, 0, offset);
ZDataType* z_matrix_ptr = (p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset);
......@@ -206,9 +214,9 @@ __global__ void
p_d_grid + lse_batch_offset,
p_ygrad_grid + c_batch_offset,
p_qgrad_grid + a_batch_offset,
p_kgrad_grid + b_batch_offset,
p_kgrad_grid + bgrad_batch_offset,
tmp_p_d0grad_grid,
p_vgrad_grid + b1_batch_offset,
p_vgrad_grid + b1grad_batch_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -217,9 +225,11 @@ __global__ void
c_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
bgrad_grid_desc_bk0_n_bk1,
d0_grid_desc_m0_n0_m1_m2_n1_m3,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
b1_grid_desc_bk0_n_bk1,
b1grad_grid_desc_bk0_n_bk1,
lse_grid_desc_m,
ygrad_grid_desc_m0_o_m1,
block_2_ctile_map,
......@@ -243,9 +253,9 @@ __global__ void
p_d_grid + lse_batch_offset,
p_ygrad_grid + c_batch_offset,
p_qgrad_grid + a_batch_offset,
p_kgrad_grid + b_batch_offset,
p_kgrad_grid + bgrad_batch_offset,
tmp_p_d0grad_grid,
p_vgrad_grid + b1_batch_offset,
p_vgrad_grid + b1grad_batch_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -254,9 +264,11 @@ __global__ void
c_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
bgrad_grid_desc_bk0_n_bk1,
d0_grid_desc_m0_n0_m1_m2_n1_m3,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
b1_grid_desc_bk0_n_bk1,
b1grad_grid_desc_bk0_n_bk1,
lse_grid_desc_m,
ygrad_grid_desc_m0_o_m1,
block_2_ctile_map,
......@@ -287,13 +299,16 @@ __global__ void
ignore = c_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = bgrad_grid_desc_bk0_n_bk1;
ignore = d0_grid_desc_m0_n0_m1_m2_n1_m3;
ignore = c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3;
ignore = b1_grid_desc_bk0_n_bk1;
ignore = b1grad_grid_desc_bk0_n_bk1;
ignore = lse_grid_desc_m;
ignore = ygrad_grid_desc_m0_o_m1;
ignore = block_2_ctile_map;
ignore = batch_count;
ignore = h_ratio;
ignore = nblock;
ignore = compute_base_ptr_of_batch;
ignore = c0_matrix_mask;
......@@ -704,6 +719,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const ZGridDesc_G_M_N& z_grid_desc_g_m_n,
const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
const CGridDesc_G_M_N& c_grid_desc_g_m_n,
const BGridDesc_G_N_K& bgrad_grid_desc_g_n_k,
const B1GridDesc_G_N_K& b1grad_grid_desc_g_n_k,
index_t BatchStrideLSE)
: a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
......@@ -711,6 +728,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
c_grid_desc_g_m_n_(c_grid_desc_g_m_n),
bgrad_grid_desc_g_n_k_(bgrad_grid_desc_g_n_k),
b1grad_grid_desc_g_n_k_(b1grad_grid_desc_g_n_k),
BatchStrideLSE_(BatchStrideLSE)
{
}
......@@ -729,6 +748,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
{
return d0_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetZBasePtr(index_t g_idx) const
{
return z_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
......@@ -749,6 +769,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
return g_idx * static_cast<long_index_t>(BatchStrideLSE_);
}
__host__ __device__ constexpr long_index_t GetBGradBasePtr(index_t g_idx) const
{
return bgrad_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetB1GradBasePtr(index_t g_idx) const
{
return b1grad_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
private:
AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_;
......@@ -756,6 +786,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
ZGridDesc_G_M_N z_grid_desc_g_m_n_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
BGridDesc_G_N_K bgrad_grid_desc_g_n_k_;
B1GridDesc_G_N_K b1grad_grid_desc_g_n_k_;
index_t BatchStrideLSE_;
};
......@@ -874,6 +906,10 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const std::vector<index_t>& c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
const std::vector<index_t>& c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
const std::vector<index_t>& lse_gs_ms_lengths,
const std::vector<index_t>& bgrad_gs_ns_ks_lengths,
const std::vector<index_t>& bgrad_gs_ns_ks_strides,
const std::vector<index_t>& b1grad_gs_gemm1ns_gemm1ks_lengths,
const std::vector<index_t>& b1grad_gs_gemm1ns_gemm1ks_strides,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_lengths,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_strides,
const std::vector<ck::index_t>&
......@@ -904,9 +940,13 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
b_grid_desc_bk0_n_bk1_{
DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
bgrad_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(
bgrad_gs_ns_ks_lengths, bgrad_gs_ns_ks_strides)},
z_grid_desc_m_n_{MakeZGridDescriptor_M_N(z_gs_ms_ns_lengths, z_gs_ms_ns_strides)},
b1_grid_desc_bk0_n_bk1_{DeviceOp::MakeVGridDescriptor_O0_N_O1(
b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
b1grad_grid_desc_bk0_n_bk1_{DeviceOp::MakeVGridDescriptor_O0_N_O1(
b1grad_gs_gemm1ns_gemm1ks_lengths, b1grad_gs_gemm1ns_gemm1ks_strides)},
y_grid_desc_m_o_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
c_gs_ms_gemm1ns_strides)},
d_y_grid_desc_m_o_{DTransform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
......@@ -927,6 +967,10 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
c_gs_ms_gemm1ns_strides)},
z_grid_desc_g_m_n_{
Transform::MakeC0GridDescriptor_G_M_N(z_gs_ms_ns_lengths, z_gs_ms_ns_strides)},
bgrad_grid_desc_g_n_k_{Transform::MakeB0GridDescriptor_G_N_K(bgrad_gs_ns_ks_lengths,
bgrad_gs_ns_ks_strides)},
b1grad_grid_desc_g_n_k_{Transform::MakeB1GridDescriptor_G_N_K(
b1grad_gs_gemm1ns_gemm1ks_lengths, b1grad_gs_gemm1ns_gemm1ks_strides)},
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(k_grid_desc_n_k_)},
d_block_2_ctile_map_{
GridwiseYDotYGrad::MakeDefaultBlock2CTileMap(d_y_grid_desc_m_o_)},
......@@ -950,6 +994,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
c_mz_gemm1nz_strides_{c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
batch_count_{c_grid_desc_g_m_n_.GetLength(I0)},
h_ratio_{c_grid_desc_g_m_n_.GetLength(I0) / b_grid_desc_g_n_k_.GetLength(I0)},
p_drop_{p_drop}
{
// TODO: implement bias addition
......@@ -979,6 +1024,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
z_grid_desc_g_m_n_,
b1_grid_desc_g_n_k_,
c_grid_desc_g_m_n_,
bgrad_grid_desc_g_n_k_,
b1grad_grid_desc_g_n_k_,
type_convert<index_t>(lse_grid_desc_m_.GetElementSpaceSize()));
seed_ = std::get<0>(seeds);
......@@ -1005,7 +1052,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
<< b_grid_desc_g_n_k_.GetLength(I1) << ", "
<< b_grid_desc_g_n_k_.GetLength(I2) << '\n';
// b_grid_desc_g_n_k_.Print();
std::cout << "b1_grid_desc_g_o_n_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
<< b1_grid_desc_g_n_k_.GetLength(I1) << ", "
<< b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
// b1_grid_desc_g_n_k_.Print();
......@@ -1018,6 +1065,17 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
std::cout << "ygrad_grid_desc_m0_o_m1_: " << ygrad_grid_desc_m0_o_m1_.GetLength(I0)
<< ", " << ygrad_grid_desc_m0_o_m1_.GetLength(I1) << ", "
<< ygrad_grid_desc_m0_o_m1_.GetLength(I2) << '\n';
std::cout << "d0_grid_desc_g_m_n_: " << d0_grid_desc_g_m_n_.GetLength(I0) << ", "
<< d0_grid_desc_g_m_n_.GetLength(I1) << ", "
<< d0_grid_desc_g_m_n_.GetLength(I2) << '\n';
std::cout << "bgrad_grid_desc_g_n_k_: " << bgrad_grid_desc_g_n_k_.GetLength(I0) << ", "
<< bgrad_grid_desc_g_n_k_.GetLength(I1) << ", "
<< bgrad_grid_desc_g_n_k_.GetLength(I2) << '\n';
// bgrad_grid_desc_g_n_k_.Print();
std::cout << "b1grad_grid_desc_g_n_k_: " << b1grad_grid_desc_g_n_k_.GetLength(I0)
<< ", " << b1grad_grid_desc_g_n_k_.GetLength(I1) << ", "
<< b1grad_grid_desc_g_n_k_.GetLength(I2) << '\n';
// b1grad_grid_desc_g_n_k_.Print();
}
// pointers
......@@ -1038,9 +1096,11 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
// tensor descriptor
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
BGridDesc_BK0_N_BK1 bgrad_grid_desc_bk0_n_bk1_;
typename GridwiseGemm::D0GridDescriptor_M0_N0_M1_M2_N1_M3 d0_grid_desc_m0_n0_m1_m2_n1_m3_;
ZGridDesc_M_N z_grid_desc_m_n_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
B1GridDesc_BK0_N_BK1 b1grad_grid_desc_bk0_n_bk1_;
YGridDesc_M_O y_grid_desc_m_o_;
DYGridDesc_M_O d_y_grid_desc_m_o_;
LSEGridDesc_M lse_grid_desc_m_;
......@@ -1055,6 +1115,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
ZGridDesc_G_M_N z_grid_desc_g_m_n_;
BGridDesc_G_N_K bgrad_grid_desc_g_n_k_;
B1GridDesc_G_N_K b1grad_grid_desc_g_n_k_;
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_;
......@@ -1083,6 +1145,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
std::vector<index_t> c_mz_gemm1nz_strides_;
index_t batch_count_;
index_t h_ratio_;
ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
float p_drop_;
......@@ -1208,13 +1271,16 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
arg.c_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.bgrad_grid_desc_bk0_n_bk1_,
arg.d0_grid_desc_m0_n0_m1_m2_n1_m3_,
arg.c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.b1grad_grid_desc_bk0_n_bk1_,
arg.lse_grid_desc_m_,
arg.ygrad_grid_desc_m0_o_m1_,
arg.block_2_ctile_map_,
arg.batch_count_,
arg.h_ratio_,
arg.block_2_ctile_map_.CalculateGridSize(arg.k_grid_desc_n_k_),
arg.compute_base_ptr_of_batch_,
arg.c0_matrix_mask_,
......@@ -1280,13 +1346,15 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
// Check if C permute dimension matches GEMM + GEMM shape
const index_t c_g = arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
const index_t b_g = arg.b_grid_desc_g_n_k_.GetLength(I0);
const index_t c_m = arg.y_grid_desc_m_o_.GetLength(I0);
const index_t c_gemm1n = arg.y_grid_desc_m_o_.GetLength(I1);
const index_t a_m = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
const index_t b1_gemm1n =
arg.b1_grid_desc_bk0_n_bk1_.GetLength(I0) * arg.b1_grid_desc_bk0_n_bk1_.GetLength(I2);
if(!(c_g == arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n))
if(!(c_g == arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n && c_g % b_g == 0 &&
b_g <= c_g))
{
return false;
}
......@@ -1380,6 +1448,10 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const std::vector<index_t>& c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
const std::vector<index_t>& c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
const std::vector<index_t>& lse_gs_ms_lengths,
const std::vector<index_t>& bgrad_gs_ns_ks_lengths,
const std::vector<index_t>& bgrad_gs_ns_ks_strides,
const std::vector<index_t>& b1grad_gs_gemm1ns_gemm1ks_lengths,
const std::vector<index_t>& b1grad_gs_gemm1ns_gemm1ks_strides,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_lengths,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_strides,
const std::vector<ck::index_t>&
......@@ -1420,6 +1492,10 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
lse_gs_ms_lengths,
bgrad_gs_ns_ks_lengths,
bgrad_gs_ns_ks_strides,
b1grad_gs_gemm1ns_gemm1ks_lengths,
b1grad_gs_gemm1ns_gemm1ks_strides,
acc0_bias_gs_ms_ns_lengths,
acc0_bias_gs_ms_ns_strides,
acc1_bias_gs_ms_gemm1ns_lengths, // acc1_bias_gs_ms_os_lengths
......@@ -1464,6 +1540,10 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const std::vector<index_t>& c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
const std::vector<index_t>& c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
const std::vector<index_t>& lse_gs_ms_lengths,
const std::vector<index_t>& bgrad_gs_ns_ks_lengths,
const std::vector<index_t>& bgrad_gs_ns_ks_strides,
const std::vector<index_t>& b1grad_gs_gemm1ns_gemm1ks_lengths,
const std::vector<index_t>& b1grad_gs_gemm1ns_gemm1ks_strides,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_lengths,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_strides,
const std::vector<ck::index_t>&
......@@ -1505,6 +1585,10 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
lse_gs_ms_lengths,
bgrad_gs_ns_ks_lengths,
bgrad_gs_ns_ks_strides,
b1grad_gs_gemm1ns_gemm1ks_lengths,
b1grad_gs_gemm1ns_gemm1ks_strides,
acc0_bias_gs_ms_ns_lengths,
acc0_bias_gs_ms_ns_strides,
acc1_bias_gs_ms_gemm1ns_lengths,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_qloop_light_v1.hpp
View file @ 104aeabc
......@@ -103,6 +103,7 @@ __global__ void
kernel_grouped_multihead_attention_backward_qloop_xdl_cshuffle_light_v1(
const void CK_CONSTANT_ADDRESS_SPACE* group_kernel_args,
const index_t group_count,
const index_t h_ratio,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op,
......@@ -141,19 +142,26 @@ __global__ void
const index_t num_blocks_per_batch = arg_ptr[group_id].num_blocks_per_batch_;
const index_t g_idx = __builtin_amdgcn_readfirstlane(
(block_id - arg_ptr[group_id].block_start_) / (Deterministic ? 1 : num_blocks_per_batch));
const index_t gkv_idx = __builtin_amdgcn_readfirstlane(g_idx / h_ratio);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(gkv_idx)));
const long_index_t z_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetZBasePtr(g_idx)));
const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(g_idx)));
arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(gkv_idx)));
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetLSEBasePtr(g_idx)));
const long_index_t bgrad_batch_offset =
__builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetBGradBasePtr(g_idx)));
const long_index_t b1grad_batch_offset =
__builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1GradBasePtr(g_idx)));
const index_t global_thread_id = get_thread_global_1d_id();
ck::philox ph(seed, global_thread_id, offset);
......@@ -168,6 +176,7 @@ __global__ void
const long_index_t d0_batch_offset =
__builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetD0BasePtr(g_idx)));
if(arg_ptr[group_id].p_d0_grid_ != nullptr)
tmp_p_d0_grid = arg_ptr[group_id].p_d0_grid_ + d0_batch_offset;
if(arg_ptr[group_id].p_d0grad_grid_)
......@@ -187,9 +196,9 @@ __global__ void
arg_ptr[group_id].p_d_grid_ + lse_batch_offset,
arg_ptr[group_id].p_ygrad_grid_ + c_batch_offset,
arg_ptr[group_id].p_qgrad_grid_ + a_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + b_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + bgrad_batch_offset,
tmp_p_d0grad_grid,
arg_ptr[group_id].p_vgrad_grid_ + b1_batch_offset,
arg_ptr[group_id].p_vgrad_grid_ + b1grad_batch_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -198,9 +207,11 @@ __global__ void
c_element_op,
arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].bgrad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].d0_grid_desc_m0_n0_m1_m2_n1_m3_,
arg_ptr[group_id].c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_,
arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].b1grad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].lse_grid_desc_m_,
arg_ptr[group_id].ygrad_grid_desc_o0_m_o1_,
arg_ptr[group_id].block_2_ctile_map_,
......@@ -225,9 +236,9 @@ __global__ void
arg_ptr[group_id].p_d_grid_ + lse_batch_offset,
arg_ptr[group_id].p_ygrad_grid_ + c_batch_offset,
arg_ptr[group_id].p_qgrad_grid_ + a_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + b_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + bgrad_batch_offset,
tmp_p_d0grad_grid,
arg_ptr[group_id].p_vgrad_grid_ + b1_batch_offset,
arg_ptr[group_id].p_vgrad_grid_ + b1grad_batch_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -236,9 +247,11 @@ __global__ void
c_element_op,
arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].bgrad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].d0_grid_desc_m0_n0_m1_m2_n1_m3_,
arg_ptr[group_id].c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_,
arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].b1grad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].lse_grid_desc_m_,
arg_ptr[group_id].ygrad_grid_desc_o0_m_o1_,
arg_ptr[group_id].block_2_ctile_map_,
......@@ -253,6 +266,7 @@ __global__ void
#else
ignore = group_kernel_args;
ignore = group_count;
ignore = h_ratio;
ignore = a_element_op;
ignore = b_element_op;
ignore = acc_element_op;
......@@ -366,6 +380,12 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
std::vector<index_t> lse_gs_ms_lengths;
std::vector<index_t> lse_gs_ms_strides;
std::vector<index_t> bgrad_gs_ns_ks_lengths;
std::vector<index_t> bgrad_gs_ns_ks_strides;
std::vector<index_t> b1grad_gs_gemm1ns_gemm1ks_lengths;
std::vector<index_t> b1grad_gs_gemm1ns_gemm1ks_strides;
std::vector<index_t> acc0_bias_gs_ms_ns_lengths;
std::vector<index_t> acc0_bias_gs_ms_ns_strides;
......@@ -576,7 +596,6 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
static auto MakeD0GridDescriptor_M_N(const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_lengths,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_strides)
{
return Transform::MakeC0GridDescriptor_M_N(acc0_bias_gs_ms_ns_lengths,
acc0_bias_gs_ms_ns_strides);
}
......@@ -585,7 +604,6 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
MakeD0GridDescriptor_G_M_N(const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_lengths,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_strides)
{
return Transform::MakeC0GridDescriptor_G_M_N(acc0_bias_gs_ms_ns_lengths,
acc0_bias_gs_ms_ns_strides);
}
......@@ -625,7 +643,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
using D0GridDesc_G_M_N = decltype(MakeD0GridDescriptor_G_M_N({}, {}));
using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
using CGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using ZGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using ZGridDesc_G_M_N = decltype(Transform::MakeC0GridDescriptor_G_M_N({}, {}));
using KGridDesc_N_K = decltype(Transform::MakeB0GridDescriptor_N_K({}, {}));
using D0GridDesc_M_N = decltype(MakeD0GridDescriptor_M_N({}, {}));
......@@ -660,6 +678,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
const ZGridDesc_G_M_N& z_grid_desc_g_m_n,
const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
const CGridDesc_G_M_N& c_grid_desc_g_m_n,
const BGridDesc_G_N_K& bgrad_grid_desc_g_n_k,
const B1GridDesc_G_N_K& b1grad_grid_desc_g_n_k,
index_t batch_stride_lse)
: a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
......@@ -667,6 +687,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
c_grid_desc_g_m_n_(c_grid_desc_g_m_n),
bgrad_grid_desc_g_n_k_(bgrad_grid_desc_g_n_k),
b1grad_grid_desc_g_n_k_(b1grad_grid_desc_g_n_k),
batch_stride_lse_(batch_stride_lse)
{
}
......@@ -706,6 +728,16 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
return g_idx * static_cast<long_index_t>(batch_stride_lse_);
}
__host__ __device__ constexpr long_index_t GetBGradBasePtr(index_t g_idx) const
{
return bgrad_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetB1GradBasePtr(index_t g_idx) const
{
return b1grad_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
private:
AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_;
......@@ -713,6 +745,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
ZGridDesc_G_M_N z_grid_desc_g_m_n_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
BGridDesc_G_N_K bgrad_grid_desc_g_n_k_;
B1GridDesc_G_N_K b1grad_grid_desc_g_n_k_;
index_t batch_stride_lse_;
};
......@@ -817,9 +851,11 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
// tensor descriptors for block/thread-wise copy
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
BGridDesc_BK0_N_BK1 bgrad_grid_desc_bk0_n_bk1_;
typename GridwiseGemm::D0GridDescriptor_M0_N0_M1_M2_N1_M3 d0_grid_desc_m0_n0_m1_m2_n1_m3_;
ZGridDesc_M_N z_grid_desc_m_n_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
B1GridDesc_BK0_N_BK1 b1grad_grid_desc_bk0_n_bk1_;
YGridDesc_M_O y_grid_desc_m_o_;
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
......@@ -861,6 +897,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
std::vector<index_t> c_mz_gemm1nz_strides_;
// for gridwise gemm check
BGridDesc_G_N_K b_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
index_t batch_count_;
......@@ -893,13 +930,15 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_drop,
index_t h_ratio,
std::tuple<unsigned long long, unsigned long long> seeds)
: a_element_op_{a_element_op},
b_element_op_{b_element_op},
acc_element_op_{acc_element_op},
b1_element_op_{b1_element_op},
c_element_op_{c_element_op},
p_dropout_{p_drop}
p_dropout_{p_drop},
h_ratio_{h_ratio}
{
seed_ = std::get<0>(seeds);
offset_ = std::get<1>(seeds);
......@@ -960,6 +999,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
const auto b_grid_desc_bk0_n_bk1 = DeviceOp::MakeBGridDescriptor_BK0_N_BK1(
problem_desc.b_gs_ns_ks_lengths, problem_desc.b_gs_ns_ks_strides);
const auto bgrad_grid_desc_bk0_n_bk1 = DeviceOp::MakeBGridDescriptor_BK0_N_BK1(
problem_desc.bgrad_gs_ns_ks_lengths, problem_desc.bgrad_gs_ns_ks_strides);
std::vector<index_t> tmp_d0_gs_ms_ns_lengths;
std::vector<index_t> tmp_d0_gs_ms_ns_strides;
......@@ -982,6 +1023,9 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
const auto b1_grid_desc_bk0_n_bk1 = DeviceOp::MakeVGridDescriptor_O0_N_O1(
problem_desc.b1_gs_gemm1ns_gemm1ks_lengths,
problem_desc.b1_gs_gemm1ns_gemm1ks_strides);
const auto b1grad_grid_desc_bk0_n_bk1 = DeviceOp::MakeVGridDescriptor_O0_N_O1(
problem_desc.b1grad_gs_gemm1ns_gemm1ks_lengths,
problem_desc.b1grad_gs_gemm1ns_gemm1ks_strides);
const auto y_grid_desc_m_o = Transform::MakeCGridDescriptor_M_N(
problem_desc.c_gs_ms_gemm1ns_lengths, problem_desc.c_gs_ms_gemm1ns_strides);
......@@ -1005,6 +1049,11 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
problem_desc.b1_gs_gemm1ns_gemm1ks_strides);
const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
problem_desc.c_gs_ms_gemm1ns_lengths, problem_desc.c_gs_ms_gemm1ns_strides);
const auto bgrad_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
problem_desc.bgrad_gs_ns_ks_lengths, problem_desc.bgrad_gs_ns_ks_strides);
const auto b1grad_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
problem_desc.b1grad_gs_gemm1ns_gemm1ks_lengths,
problem_desc.b1grad_gs_gemm1ns_gemm1ks_strides);
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3;
const index_t BlockStart = grid_size_;
......@@ -1027,6 +1076,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
z_grid_desc_g_m_n,
b1_grid_desc_g_n_k,
c_grid_desc_g_m_n,
bgrad_grid_desc_g_n_k,
b1grad_grid_desc_g_n_k,
type_convert<index_t>(lse_grid_desc_m.GetElementSpaceSize()));
// C0 mask
......@@ -1073,9 +1124,11 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
p_vgrad_grid,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
bgrad_grid_desc_bk0_n_bk1,
d0_grid_desc_m0_n0_m1_m2_n1_m3,
z_grid_desc_m_n,
b1_grid_desc_bk0_n_bk1,
b1grad_grid_desc_bk0_n_bk1,
y_grid_desc_m_o,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
lse_grid_desc_m,
......@@ -1119,6 +1172,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
problem_desc.b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
{problem_desc.c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
problem_desc.c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
b_grid_desc_g_n_k,
c_grid_desc_g_m_n,
batch_count,
d0_n_length_stride});
......@@ -1145,6 +1199,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
index_t grid_size_;
index_t group_count_;
index_t h_ratio_;
std::vector<GroupKernelArg> group_kernel_args_;
std::vector<GroupDeviceArg> group_device_args_;
......@@ -1224,6 +1279,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
0,
cast_pointer_to_constant_address_space(arg.p_workspace_),
arg.group_count_,
arg.h_ratio_,
arg.a_element_op_,
arg.b_element_op_,
arg.acc_element_op_,
......@@ -1292,13 +1348,15 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
const auto& device_arg = arg.group_device_args_[i];
// Check if C permute dimension matches GEMM + GEMM shape
const index_t c_g = device_arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
const index_t b_g = device_arg.b_grid_desc_g_n_k_.GetLength(I0);
const index_t c_m = kernel_arg.y_grid_desc_m_o_.GetLength(I0);
const index_t c_gemm1n = kernel_arg.y_grid_desc_m_o_.GetLength(I1);
const index_t a_m = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
const index_t b1_gemm1n = kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I0) *
kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I2);
if(!(c_g == device_arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n))
if(!(c_g == device_arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n &&
c_g % b_g == 0 && c_g / b_g == arg.h_ratio_))
{
return false;
}
......@@ -1395,6 +1453,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_drop,
index_t h_ratio,
std::tuple<unsigned long long, unsigned long long> seeds)
{
return Argument{p_As,
......@@ -1419,6 +1478,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
b1_element_op,
c_element_op,
p_drop,
h_ratio,
seeds};
}
......@@ -1449,6 +1509,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_drop,
index_t h_ratio,
std::tuple<unsigned long long, unsigned long long> seeds) // override
{
return std::make_unique<Argument>(p_As,
......@@ -1473,6 +1534,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
b1_element_op,
c_element_op,
p_drop,
h_ratio,
seeds);
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_qloop_light_v2.hpp
View file @ 104aeabc
......@@ -102,6 +102,7 @@ __global__ void
kernel_grouped_multihead_attention_backward_qloop_xdl_cshuffle_light_v2(
const void CK_CONSTANT_ADDRESS_SPACE* group_kernel_args,
const index_t group_count,
const index_t h_ratio,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op,
......@@ -140,19 +141,26 @@ __global__ void
const index_t num_blocks_per_batch = arg_ptr[group_id].num_blocks_per_batch_;
const index_t g_idx = __builtin_amdgcn_readfirstlane(
(block_id - arg_ptr[group_id].block_start_) / (Deterministic ? 1 : num_blocks_per_batch));
const index_t gkv_idx = __builtin_amdgcn_readfirstlane(g_idx / h_ratio);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(gkv_idx)));
const long_index_t z_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetZBasePtr(g_idx)));
const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(g_idx)));
arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(gkv_idx)));
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetLSEBasePtr(g_idx)));
const long_index_t bgrad_batch_offset =
__builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetBGradBasePtr(g_idx)));
const long_index_t b1grad_batch_offset =
__builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1GradBasePtr(g_idx)));
const index_t global_thread_id = get_thread_global_1d_id();
ck::philox ph(seed, global_thread_id, offset);
......@@ -166,7 +174,6 @@ __global__ void
const long_index_t d0_batch_offset =
__builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetD0BasePtr(g_idx)));
if(arg_ptr[group_id].p_d0_grid_ != nullptr)
tmp_p_d0_grid = arg_ptr[group_id].p_d0_grid_ + d0_batch_offset;
if(arg_ptr[group_id].p_d0grad_grid_)
......@@ -187,9 +194,9 @@ __global__ void
arg_ptr[group_id].p_d_grid_ + lse_batch_offset,
arg_ptr[group_id].p_ygrad_grid_ + c_batch_offset,
arg_ptr[group_id].p_qgrad_grid_ + a_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + b_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + bgrad_batch_offset,
tmp_p_d0grad_grid,
arg_ptr[group_id].p_vgrad_grid_ + b1_batch_offset,
arg_ptr[group_id].p_vgrad_grid_ + b1grad_batch_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -198,9 +205,11 @@ __global__ void
c_element_op,
arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].bgrad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].d0_grid_desc_m0_n0_m1_m2_n1_m3_,
arg_ptr[group_id].c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_,
arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].b1grad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].lse_grid_desc_m_,
arg_ptr[group_id].ygrad_grid_desc_m0_o_m1_,
arg_ptr[group_id].block_2_ctile_map_,
......@@ -225,9 +234,9 @@ __global__ void
arg_ptr[group_id].p_d_grid_ + lse_batch_offset,
arg_ptr[group_id].p_ygrad_grid_ + c_batch_offset,
arg_ptr[group_id].p_qgrad_grid_ + a_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + b_batch_offset,
arg_ptr[group_id].p_kgrad_grid_ + bgrad_batch_offset,
tmp_p_d0grad_grid,
arg_ptr[group_id].p_vgrad_grid_ + b1_batch_offset,
arg_ptr[group_id].p_vgrad_grid_ + b1grad_batch_offset,
p_shared,
a_element_op,
b_element_op,
......@@ -236,9 +245,11 @@ __global__ void
c_element_op,
arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].bgrad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].d0_grid_desc_m0_n0_m1_m2_n1_m3_,
arg_ptr[group_id].c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_,
arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].b1grad_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].lse_grid_desc_m_,
arg_ptr[group_id].ygrad_grid_desc_m0_o_m1_,
arg_ptr[group_id].block_2_ctile_map_,
......@@ -253,6 +264,7 @@ __global__ void
#else
ignore = group_kernel_args;
ignore = group_count;
ignore = h_ratio;
ignore = a_element_op;
ignore = b_element_op;
ignore = acc_element_op;
......@@ -373,6 +385,12 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
std::vector<index_t> lse_gs_ms_lengths;
std::vector<index_t> lse_gs_ms_strides;
std::vector<index_t> bgrad_gs_ns_ks_lengths;
std::vector<index_t> bgrad_gs_ns_ks_strides;
std::vector<index_t> b1grad_gs_gemm1ns_gemm1ks_lengths;
std::vector<index_t> b1grad_gs_gemm1ns_gemm1ks_strides;
std::vector<index_t> acc0_bias_gs_ms_ns_lengths;
std::vector<index_t> acc0_bias_gs_ms_ns_strides;
......@@ -639,7 +657,6 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
static auto MakeD0GridDescriptor_M_N(const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_lengths,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_strides)
{
return Transform::MakeC0GridDescriptor_M_N(acc0_bias_gs_ms_ns_lengths,
acc0_bias_gs_ms_ns_strides);
}
......@@ -648,7 +665,6 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
MakeD0GridDescriptor_G_M_N(const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_lengths,
const std::vector<ck::index_t>& acc0_bias_gs_ms_ns_strides)
{
return Transform::MakeC0GridDescriptor_G_M_N(acc0_bias_gs_ms_ns_lengths,
acc0_bias_gs_ms_ns_strides);
}
......@@ -723,6 +739,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const ZGridDesc_G_M_N& z_grid_desc_g_m_n,
const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
const CGridDesc_G_M_N& c_grid_desc_g_m_n,
const BGridDesc_G_N_K& bgrad_grid_desc_g_n_k,
const B1GridDesc_G_N_K& b1grad_grid_desc_g_n_k,
index_t BatchStrideLSE)
: a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
......@@ -730,6 +748,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
c_grid_desc_g_m_n_(c_grid_desc_g_m_n),
bgrad_grid_desc_g_n_k_(bgrad_grid_desc_g_n_k),
b1grad_grid_desc_g_n_k_(b1grad_grid_desc_g_n_k),
BatchStrideLSE_(BatchStrideLSE)
{
}
......@@ -769,6 +789,16 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
return g_idx * static_cast<long_index_t>(BatchStrideLSE_);
}
__host__ __device__ constexpr long_index_t GetBGradBasePtr(index_t g_idx) const
{
return bgrad_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetB1GradBasePtr(index_t g_idx) const
{
return b1grad_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
private:
AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_;
......@@ -776,6 +806,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
ZGridDesc_G_M_N z_grid_desc_g_m_n_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
BGridDesc_G_N_K bgrad_grid_desc_g_n_k_;
B1GridDesc_G_N_K b1grad_grid_desc_g_n_k_;
index_t BatchStrideLSE_;
};
......@@ -888,9 +920,11 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
// tensor descriptors for block/thread-wise copy
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
BGridDesc_BK0_N_BK1 bgrad_grid_desc_bk0_n_bk1_;
typename GridwiseGemm::D0GridDescriptor_M0_N0_M1_M2_N1_M3 d0_grid_desc_m0_n0_m1_m2_n1_m3_;
ZGridDesc_M_N z_grid_desc_m_n_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
B1GridDesc_BK0_N_BK1 b1grad_grid_desc_bk0_n_bk1_;
YGridDesc_M_O y_grid_desc_m_o_;
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
......@@ -932,6 +966,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
std::vector<index_t> c_mz_gemm1nz_strides_;
// for gridwise gemm check
BGridDesc_G_N_K b_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
index_t batch_count_;
......@@ -964,13 +999,15 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_drop,
index_t h_ratio,
std::tuple<unsigned long long, unsigned long long> seeds)
: a_element_op_{a_element_op},
b_element_op_{b_element_op},
acc_element_op_{acc_element_op},
b1_element_op_{b1_element_op},
c_element_op_{c_element_op},
p_dropout_{p_drop}
p_dropout_{p_drop},
h_ratio_{h_ratio}
{
seed_ = std::get<0>(seeds);
offset_ = std::get<1>(seeds);
......@@ -1031,6 +1068,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
const auto b_grid_desc_bk0_n_bk1 = DeviceOp::MakeBGridDescriptor_BK0_N_BK1(
problem_desc.b_gs_ns_ks_lengths, problem_desc.b_gs_ns_ks_strides);
const auto bgrad_grid_desc_bk0_n_bk1 = DeviceOp::MakeBGridDescriptor_BK0_N_BK1(
problem_desc.bgrad_gs_ns_ks_lengths, problem_desc.bgrad_gs_ns_ks_strides);
std::vector<index_t> tmp_d0_gs_ms_ns_lengths;
std::vector<index_t> tmp_d0_gs_ms_ns_strides;
......@@ -1053,6 +1092,9 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const auto b1_grid_desc_bk0_n_bk1 = DeviceOp::MakeVGridDescriptor_O0_N_O1(
problem_desc.b1_gs_gemm1ns_gemm1ks_lengths,
problem_desc.b1_gs_gemm1ns_gemm1ks_strides);
const auto b1grad_grid_desc_bk0_n_bk1 = DeviceOp::MakeVGridDescriptor_O0_N_O1(
problem_desc.b1grad_gs_gemm1ns_gemm1ks_lengths,
problem_desc.b1grad_gs_gemm1ns_gemm1ks_strides);
const auto y_grid_desc_m_o = Transform::MakeCGridDescriptor_M_N(
problem_desc.c_gs_ms_gemm1ns_lengths, problem_desc.c_gs_ms_gemm1ns_strides);
......@@ -1076,6 +1118,11 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
problem_desc.b1_gs_gemm1ns_gemm1ks_strides);
const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
problem_desc.c_gs_ms_gemm1ns_lengths, problem_desc.c_gs_ms_gemm1ns_strides);
const auto bgrad_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
problem_desc.bgrad_gs_ns_ks_lengths, problem_desc.bgrad_gs_ns_ks_strides);
const auto b1grad_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
problem_desc.b1grad_gs_gemm1ns_gemm1ks_lengths,
problem_desc.b1grad_gs_gemm1ns_gemm1ks_strides);
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3;
const index_t BlockStart = grid_size_;
......@@ -1098,6 +1145,8 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
z_grid_desc_g_m_n,
b1_grid_desc_g_n_k,
c_grid_desc_g_m_n,
bgrad_grid_desc_g_n_k,
b1grad_grid_desc_g_n_k,
type_convert<index_t>(lse_grid_desc_m.GetElementSpaceSize()));
// C0 mask
......@@ -1144,9 +1193,11 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
p_vgrad_grid,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
bgrad_grid_desc_bk0_n_bk1,
d0_grid_desc_m0_n0_m1_m2_n1_m3,
z_grid_desc_m_n,
b1_grid_desc_bk0_n_bk1,
b1grad_grid_desc_bk0_n_bk1,
y_grid_desc_m_o,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
lse_grid_desc_m,
......@@ -1190,6 +1241,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
problem_desc.b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
{problem_desc.c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
problem_desc.c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
b_grid_desc_g_n_k,
c_grid_desc_g_m_n,
batch_count,
d0_n_length_stride});
......@@ -1216,6 +1268,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
index_t grid_size_;
index_t group_count_;
index_t h_ratio_;
std::vector<GroupKernelArg> group_kernel_args_;
std::vector<GroupDeviceArg> group_device_args_;
......@@ -1294,6 +1347,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
0,
cast_pointer_to_constant_address_space(arg.p_workspace_),
arg.group_count_,
arg.h_ratio_,
arg.a_element_op_,
arg.b_element_op_,
arg.acc_element_op_,
......@@ -1362,13 +1416,15 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const auto& device_arg = arg.group_device_args_[i];
// Check if C permute dimension matches GEMM + GEMM shape
const index_t c_g = device_arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
const index_t b_g = device_arg.b_grid_desc_g_n_k_.GetLength(I0);
const index_t c_m = kernel_arg.y_grid_desc_m_o_.GetLength(I0);
const index_t c_gemm1n = kernel_arg.y_grid_desc_m_o_.GetLength(I1);
const index_t a_m = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
const index_t b1_gemm1n = kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I0) *
kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I2);
if(!(c_g == device_arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n))
if(!(c_g == device_arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n &&
c_g % b_g == 0 && c_g / b_g == arg.h_ratio_))
{
return false;
}
......@@ -1471,6 +1527,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_drop,
index_t h_ratio,
std::tuple<unsigned long long, unsigned long long> seeds)
{
return Argument{p_As,
......@@ -1495,6 +1552,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
b1_element_op,
c_element_op,
p_drop,
h_ratio,
seeds};
}
......@@ -1525,6 +1583,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_drop,
index_t h_ratio,
std::tuple<unsigned long long, unsigned long long> seeds) // override
{
return std::make_unique<Argument>(p_As,
......@@ -1549,6 +1608,7 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
b1_element_op,
c_element_op,
p_drop,
h_ratio,
seeds);
}
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_mha_bwd_xdl_cshuffle_qloop_b2t_light_v1.hpp
View file @ 104aeabc
......@@ -1440,10 +1440,12 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
const CElementwiseOperation& c_element_op,
const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1,
const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1,
const KGridDesc_K0_N_K1& kgrad_grid_desc_k0_n_k1,
const D0GridDescriptor_M0_N0_M1_M2_N1_M3& d0_grid_desc_m0_n0_m1_m2_n1_m3,
const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3&
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
const VGridDesc_O0_N_O1& v_grid_desc_o0_n_o1,
const VGridDesc_O0_N_O1& vgrad_grid_desc_o0_n_o1,
const LSEGridDesc_M& lse_grid_desc_m,
const YGradGridDesc_O0_M_O1& ygrad_grid_desc_o0_m_o1,
const Block2CTileMap& block_2_ctile_map,
......@@ -1474,11 +1476,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
const auto ygrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ygrad_grid, ygrad_grid_desc_o0_m_o1.GetElementSpaceSize());
auto vgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_vgrad_grid, v_grid_desc_o0_n_o1.GetElementSpaceSize());
p_vgrad_grid, vgrad_grid_desc_o0_n_o1.GetElementSpaceSize());
auto qgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_qgrad_grid, q_grid_desc_k0_m_k1.GetElementSpaceSize());
auto kgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_kgrad_grid, k_grid_desc_k0_n_k1.GetElementSpaceSize());
p_kgrad_grid, kgrad_grid_desc_k0_n_k1.GetElementSpaceSize());
// divide block work by [N, K]
const auto block_work_idx =
......@@ -1628,7 +1630,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
// dV: transform input and output tensor descriptors
auto vgrad_grid_desc_nblock_nperblock_oblock_operblock =
MakeVGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(v_grid_desc_o0_n_o1);
MakeVGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(vgrad_grid_desc_o0_n_o1);
// dK: A matrix blockwise copy
auto kgrad_gemm_tile_sgrad_blockwise_copy =
......@@ -1657,7 +1659,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V1
// dK: transform input and output tensor descriptors
auto kgrad_grid_desc_nblock_nperblock_oblock_operblock =
MakeKGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(k_grid_desc_k0_n_k1);
MakeKGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(kgrad_grid_desc_k0_n_k1);
//
// set up dQ Gemm (type 3 crr)
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_mha_bwd_xdl_cshuffle_qloop_b2t_light_v2.hpp
View file @ 104aeabc
......@@ -1531,10 +1531,12 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const CElementwiseOperation& c_element_op,
const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1,
const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1,
const KGridDesc_K0_N_K1& kgrad_grid_desc_k0_n_k1,
const D0GridDescriptor_M0_N0_M1_M2_N1_M3& d0_grid_desc_m0_n0_m1_m2_n1_m3,
const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3&
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
const VGridDesc_O0_N_O1& v_grid_desc_o0_n_o1,
const VGridDesc_O0_N_O1& vgrad_grid_desc_o0_n_o1,
const LSEGridDesc_M& lse_grid_desc_m,
const YGradGridDesc_M0_O_M1& ygrad_grid_desc_m0_o_m1,
const Block2CTileMap& block_2_ctile_map,
......@@ -1565,11 +1567,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
const auto ygrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ygrad_grid, ygrad_grid_desc_m0_o_m1.GetElementSpaceSize());
auto vgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_vgrad_grid, v_grid_desc_o0_n_o1.GetElementSpaceSize());
p_vgrad_grid, vgrad_grid_desc_o0_n_o1.GetElementSpaceSize());
auto qgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_qgrad_grid, q_grid_desc_k0_m_k1.GetElementSpaceSize());
auto kgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_kgrad_grid, k_grid_desc_k0_n_k1.GetElementSpaceSize());
p_kgrad_grid, kgrad_grid_desc_k0_n_k1.GetElementSpaceSize());
// divide block work by [N, K]
const auto block_work_idx =
......@@ -1742,7 +1744,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
// dV: transform input and output tensor descriptors
auto vgrad_grid_desc_nblock_nperblock_oblock_operblock =
MakeVGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(v_grid_desc_o0_n_o1);
MakeVGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(vgrad_grid_desc_o0_n_o1);
// dK: transform input and output tensor descriptors
const auto q_grid_desc_m0_k_m1 =
......@@ -1775,7 +1777,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_Light_V2
// dK: transform input and output tensor descriptors
auto kgrad_grid_desc_nblock_nperblock_oblock_operblock =
MakeKGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(k_grid_desc_k0_n_k1);
MakeKGradGridDesc_NBlock_NPerBlock_OBlock_OPerBlock(kgrad_grid_desc_k0_n_k1);
//
// set up dQ Gemm (type 3 crr)
......
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