derek.holloway
7502bf64a4
CodeQL / Analyze (actions) (push) Canceled after 0s
CodeQL / Analyze (python) (push) Canceled after 0s
pre-commit / pre-commit (push) Canceled after 0s
Multi-Arch CI / setup (push) Canceled after 0s
Publish build_manylinux_x86_64 images / publish_build_manylinux_x86_64 (push) Canceled after 0s
Setup (Multi-Arch) / setup (push) Canceled after 0s
Multi-Arch CI / Linux::${{ fromJSON(needs.setup.outputs.linux_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubi10 image / publish_no_rocm_image_ubi10 (push) Canceled after 0s
Multi-Arch CI / Windows::${{ fromJSON(needs.setup.outputs.windows_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Multi-Arch CI / Manifest Diff (push) Canceled after 0s
TheRock Manifest Diff Report / Generate Manifest Diff Report (push) Canceled after 0s
Multi-Arch CI / CI Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_ocl_rt images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_rocgdb images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Unit Tests / Unit Tests :: ubuntu-24.04 (push) Canceled after 0s
Unit Tests / Unit Tests :: windows-2022 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_media (push) Canceled after 0s
Unit Tests / Unit Tests Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_openmpi (push) Canceled after 0s
1957 lines
83 KiB
C++
1957 lines
83 KiB
C++
/*******************************************************************************
|
|
*
|
|
* MIT License
|
|
*
|
|
* Copyright (c) 2025 Advanced Micro Devices, Inc.
|
|
*
|
|
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
* of this software and associated documentation files (the "Software"), to deal
|
|
* in the Software without restriction, including without limitation the rights
|
|
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
* copies of the Software, and to permit persons to whom the Software is
|
|
* furnished to do so, subject to the following conditions:
|
|
*
|
|
* The above copyright notice and this permission notice shall be included in all
|
|
* copies or substantial portions of the Software.
|
|
*
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
* SOFTWARE.
|
|
*
|
|
*******************************************************************************/
|
|
#ifndef GUARD_MIOPEN_GRU_VERIFY_GEMM_HPP
|
|
#define GUARD_MIOPEN_GRU_VERIFY_GEMM_HPP
|
|
|
|
#include "dropout_gpu_emulator.hpp"
|
|
|
|
#include <../test/rnn_util.hpp>
|
|
|
|
#include <algorithm>
|
|
#include <cassert>
|
|
#include <math.h>
|
|
|
|
template <typename Tgpu, typename Tref>
|
|
void RunGRUForwardGEMMCPUVerify(miopenHandle_t handle,
|
|
std::vector<Tgpu>& in,
|
|
std::vector<Tgpu>& wei, // [ input_state_weight_trans
|
|
// hidden_state_weight0_trans input1_trans
|
|
// hidden1_trans ... output_weight;
|
|
// bidirectional reversed weights ]
|
|
std::vector<Tref>& hy_host, // current/final hidden state
|
|
std::vector<Tgpu>& hx, // initial hidden state
|
|
std::vector<Tref>& out_host,
|
|
std::vector<int>& in_n, // input batch size
|
|
int in_h, // input data length
|
|
int seqLength, // Number of iterations to unroll over
|
|
bool bidirection, // whether using bidirectional net
|
|
bool biased, // whether using bias
|
|
int hy_d, // 1 by numlayer (number of stacks of hidden layers) for
|
|
// unidirection, 2 by numlayer for bidirection
|
|
int hy_n, // equal to input batch size in_n[0]
|
|
int hy_h, // hidden state number
|
|
int out_h, // 1 by hy_h related function for unidirection, 2 by hy_h
|
|
// related function for bidirection
|
|
int inputMode,
|
|
std::vector<Tref>& rsvspace_host,
|
|
bool use_dropout,
|
|
miopenDropoutDescriptor_t dropoutDesc,
|
|
bool hx_is_null = false)
|
|
{
|
|
int batch_n = sumvc(in_n);
|
|
|
|
int numlayer = bidirection ? hy_d / 2 : hy_d;
|
|
int bacc, baccbi; // accumulation of batch
|
|
int bi = bidirection ? 2 : 1;
|
|
|
|
int in_stride = in_h;
|
|
int out_stride = out_h;
|
|
int wei_stride = bi * 3 * hy_h;
|
|
int hy_stride = bi * 4 * hy_h;
|
|
int h_stride = bi * hy_h;
|
|
int uni_stride = hy_h;
|
|
int bi_stride = hy_h * bi;
|
|
|
|
std::vector<Tref> hid_state(numlayer * batch_n * hy_stride * 2, static_cast<Tref>(0));
|
|
|
|
Tref* out_state = new Tref[batch_n * out_h];
|
|
memset(out_state, 0, batch_n * out_h * sizeof(Tref));
|
|
|
|
// initial input
|
|
Tref* in_state = new Tref[batch_n * in_h];
|
|
for(int h = 0; h < batch_n; h++)
|
|
{
|
|
for(int w = 0; w < in_h; w++)
|
|
{
|
|
in_state[h * in_stride + w] = in[h * in_stride + w];
|
|
}
|
|
}
|
|
|
|
// initial hidden states
|
|
Tref* hy_state = new Tref[hy_d * hy_n * hy_h];
|
|
memset(hy_state, 0, hy_d * hy_n * hy_h * sizeof(Tref));
|
|
Tref* hx_state = new Tref[hy_d * hy_n * hy_h];
|
|
for(int h = 0; h < hy_d * hy_n * hy_h; h++)
|
|
{
|
|
hx_state[h] = hx[h];
|
|
}
|
|
|
|
if(inputMode == 1)
|
|
{
|
|
if(in_h != hy_h)
|
|
{
|
|
printf("Verification cannot be completed: The input tensor size must equal to the "
|
|
"hidden state size of the network in SKIP_INPUT mode!\n");
|
|
return;
|
|
}
|
|
in_h = 0;
|
|
}
|
|
|
|
int wei_shift_bias = (in_h + hy_h + (bi * hy_h + hy_h) * (numlayer - 1)) * wei_stride;
|
|
int wei_len = wei_shift_bias;
|
|
if(biased)
|
|
{
|
|
int in_bias = 2;
|
|
wei_len += (in_bias + (numlayer - 1) * 2) * wei_stride;
|
|
}
|
|
|
|
// initial weights
|
|
Tref* wei_state = new Tref[wei_len];
|
|
for(int h = 0; h < wei_len; h++)
|
|
{
|
|
wei_state[h] = wei[h];
|
|
}
|
|
|
|
// initial dropoput
|
|
std::vector<rocrand_state_xorwow> dropout_states_host;
|
|
std::vector<unsigned char> dropout_reservespace_host;
|
|
std::vector<Tref> dropout_hid_state;
|
|
miopenTensorDescriptor_t dropout_inputTensor{}, dropout_outputTensor{};
|
|
if(use_dropout)
|
|
{
|
|
size_t statesSizeInBytes = 0;
|
|
miopenDropoutGetStatesSize(handle, &statesSizeInBytes);
|
|
size_t states_size = statesSizeInBytes / sizeof(rocrand_state_xorwow);
|
|
dropout_states_host = std::vector<rocrand_state_xorwow>(states_size);
|
|
InitKernelStateEmulator(dropout_states_host, dropoutDesc);
|
|
|
|
std::array<int, 2> drop_in_len = {{batch_n, hy_h * bi}};
|
|
std::array<int, 2> drop_in_str = {{hy_stride, 1}};
|
|
std::array<int, 2> drop_out_str = {{hy_h * bi, 1}};
|
|
miopenCreateTensorDescriptor(&dropout_inputTensor);
|
|
miopenCreateTensorDescriptor(&dropout_outputTensor);
|
|
miopenSetTensorDescriptor(
|
|
dropout_inputTensor, miopenFloat, 2, drop_in_len.data(), drop_in_str.data());
|
|
miopenSetTensorDescriptor(
|
|
dropout_outputTensor, miopenFloat, 2, drop_in_len.data(), drop_out_str.data());
|
|
|
|
size_t reserveSpaceSizeInBytes = 0;
|
|
miopenDropoutGetReserveSpaceSize(dropout_inputTensor, &reserveSpaceSizeInBytes);
|
|
size_t reserve_size = reserveSpaceSizeInBytes / sizeof(unsigned char);
|
|
dropout_reservespace_host = std::vector<unsigned char>(reserve_size * (numlayer - 1),
|
|
static_cast<unsigned char>(1));
|
|
|
|
dropout_hid_state =
|
|
std::vector<Tref>((numlayer - 1) * batch_n * hy_h * bi, static_cast<Tref>(0));
|
|
}
|
|
|
|
// forward emulator
|
|
for(int li = 0; li < numlayer; li++)
|
|
{
|
|
int hid_shift = li * batch_n * hy_stride;
|
|
int hx_shift = li * in_n[0] * h_stride;
|
|
int wei_shift_bias_temp = wei_shift_bias + li * 2 * wei_stride;
|
|
|
|
// from input
|
|
if(li == 0)
|
|
{
|
|
if(inputMode == 1)
|
|
{
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
for(int gi = 0; gi < 3; gi++)
|
|
{
|
|
hid_state[hid_shift + bs * hy_stride + gi * hy_h + h] +=
|
|
in_state[bs * in_stride + h];
|
|
if(bidirection)
|
|
{
|
|
hid_state[hid_shift + bs * hy_stride + (gi + 3) * hy_h + h] +=
|
|
in_state[bs * in_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
// from bias
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < wei_stride; h++)
|
|
{
|
|
hid_state[hid_shift + bs * hy_stride + h] += wei[wei_shift_bias + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
gemm_cpu<Tref>(const_cast<Tref*>(in_state),
|
|
in_h,
|
|
batch_n,
|
|
in_stride,
|
|
false,
|
|
const_cast<Tref*>(wei_state),
|
|
in_h,
|
|
hy_h * bi * 3,
|
|
in_stride,
|
|
true,
|
|
&hid_state[hid_shift],
|
|
hy_h * bi * 3,
|
|
batch_n,
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
// from bias
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < wei_stride; h++)
|
|
{
|
|
hid_state[hid_shift + bs * hy_stride + h] += wei[wei_shift_bias + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
int wei_shift = (in_h + hy_h) * wei_stride + (li - 1) * (bi * hy_h + hy_h) * wei_stride;
|
|
int prelayer_shift = (li - 1) * batch_n * hy_stride + bi * 3 * hy_h;
|
|
if(use_dropout)
|
|
{
|
|
auto dropout_states_tmp = dropout_states_host;
|
|
size_t drop_out_offset = (li - 1) * batch_n * hy_h * bi;
|
|
|
|
RunDropoutForwardEmulator<Tref>(handle,
|
|
dropoutDesc,
|
|
dropout_inputTensor,
|
|
dropout_inputTensor,
|
|
hid_state,
|
|
dropout_outputTensor,
|
|
dropout_hid_state,
|
|
dropout_reservespace_host,
|
|
dropout_states_tmp,
|
|
prelayer_shift,
|
|
drop_out_offset,
|
|
drop_out_offset);
|
|
|
|
prelayer_shift = drop_out_offset;
|
|
}
|
|
|
|
gemm_cpu<Tref>(use_dropout ? &dropout_hid_state[prelayer_shift]
|
|
: &hid_state[prelayer_shift],
|
|
hy_h * bi,
|
|
batch_n,
|
|
use_dropout ? hy_h * bi : hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift]),
|
|
hy_h * bi,
|
|
hy_h * bi * 3,
|
|
bi_stride,
|
|
true,
|
|
&hid_state[hid_shift],
|
|
hy_h * bi * 3,
|
|
batch_n,
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
// from bias
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < wei_stride; h++)
|
|
{
|
|
hid_state[hid_shift + bs * hy_stride + h] += wei[wei_shift_bias_temp + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
// from hidden state
|
|
bacc = 0;
|
|
baccbi = batch_n;
|
|
for(int ti = 0; ti < seqLength; ti++)
|
|
{
|
|
baccbi -= in_n[seqLength - 1 - ti];
|
|
int wei_shift = in_h * wei_stride + li * (bi * hy_h + hy_h) * wei_stride;
|
|
int pretime_shift;
|
|
|
|
if(ti == 0)
|
|
{
|
|
if(!hx_is_null)
|
|
{
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&hx_state[hx_shift]),
|
|
hy_h,
|
|
in_n[ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + bacc * hy_stride],
|
|
hy_h * 2,
|
|
in_n[ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
for(int gi = 0; gi < 2; gi++)
|
|
{
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + gi * hy_h +
|
|
h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + gi * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&hx_state[hx_shift]),
|
|
hy_h,
|
|
in_n[ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 2 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + bacc * hy_stride + bi * 3 * hy_h],
|
|
hy_h,
|
|
in_n[ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h +
|
|
h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + 2 * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
|
|
if(bidirection)
|
|
{
|
|
gemm_cpu<Tref>(
|
|
const_cast<Tref*>(&hx_state[hx_shift + hy_n * hy_h]),
|
|
hy_h,
|
|
in_n[seqLength - 1 - ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 3 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + baccbi * hy_stride + 3 * hy_h],
|
|
hy_h * 2,
|
|
in_n[seqLength - 1 - ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[seqLength - 1 - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
for(int gi = 0; gi < 2; gi++)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
(3 + gi) * hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + (3 + gi) * hy_h +
|
|
h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(
|
|
const_cast<Tref*>(&hx_state[hx_shift + hy_n * hy_h]),
|
|
hy_h,
|
|
in_n[seqLength - 1 - ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 5 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + baccbi * hy_stride + bi * 3 * hy_h + hy_h],
|
|
hy_h,
|
|
in_n[seqLength - 1 - ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[seqLength - 1 - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
bi * 3 * hy_h + hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + 5 * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&hy_state[hx_shift]),
|
|
hy_h,
|
|
in_n[ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + bacc * hy_stride],
|
|
hy_h * 2,
|
|
in_n[ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
for(int gi = 0; gi < 2; gi++)
|
|
{
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + gi * hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + gi * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&hy_state[hx_shift]),
|
|
hy_h,
|
|
in_n[ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 2 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + bacc * hy_stride + bi * 3 * hy_h],
|
|
hy_h,
|
|
in_n[ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + 2 * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
|
|
if(bidirection)
|
|
{
|
|
|
|
if(!hx_is_null && in_n.at(seqLength - 1 - ti) > in_n.at(seqLength - ti))
|
|
{
|
|
gemm_cpu<Tref>(
|
|
const_cast<Tref*>(
|
|
&hx_state[hx_shift + hy_n * hy_h + in_n.at(seqLength - ti) * hy_h]),
|
|
hy_h,
|
|
(in_n.at(seqLength - 1 - ti) - in_n.at(seqLength - ti)),
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 3 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + (baccbi + in_n.at(seqLength - ti)) * hy_stride +
|
|
3 * hy_h],
|
|
hy_h * 2,
|
|
(in_n.at(seqLength - 1 - ti) - in_n.at(seqLength - ti)),
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = in_n.at(seqLength - ti); bs < in_n.at(seqLength - 1 - ti);
|
|
bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
for(int gi = 0; gi < 2; gi++)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
(3 + gi) * hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + (3 + gi) * hy_h +
|
|
h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(
|
|
const_cast<Tref*>(
|
|
&hx_state[hx_shift + hy_n * hy_h + in_n.at(seqLength - ti) * hy_h]),
|
|
hy_h,
|
|
(in_n.at(seqLength - 1 - ti) - in_n.at(seqLength - ti)),
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 5 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + (baccbi + in_n.at(seqLength - ti)) * hy_stride +
|
|
bi * 3 * hy_h + hy_h],
|
|
hy_h,
|
|
(in_n.at(seqLength - 1 - ti) - in_n.at(seqLength - ti)),
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = in_n.at(seqLength - ti); bs < in_n.at(seqLength - 1 - ti);
|
|
bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
bi * 3 * hy_h + hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + 5 * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&hy_state[hx_shift + hy_n * hy_h]),
|
|
hy_h,
|
|
in_n[seqLength - ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 3 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + baccbi * hy_stride + 3 * hy_h],
|
|
hy_h * 2,
|
|
in_n[seqLength - ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[seqLength - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
for(int gi = 0; gi < 2; gi++)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
(3 + gi) * hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + (3 + gi) * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(
|
|
const_cast<Tref*>(&hy_state[hx_shift + hy_n * hy_h]),
|
|
hy_h,
|
|
in_n[seqLength - ti],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift + 5 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
true,
|
|
&hid_state[hid_shift + baccbi * hy_stride + bi * 3 * hy_h + hy_h],
|
|
hy_h,
|
|
in_n[seqLength - ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
for(int bs = 0; bs < in_n[seqLength - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h] +=
|
|
wei[wei_shift_bias_temp + wei_stride + 5 * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h];
|
|
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h] +=
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + hy_h + h], 2) *
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h];
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] = 0;
|
|
|
|
if(ti == 0)
|
|
{
|
|
if(!hx_is_null)
|
|
{
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] +=
|
|
((1 - activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + h],
|
|
2)) *
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride +
|
|
2 * hy_h + h],
|
|
1) +
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + h], 2) *
|
|
hx[hx_shift + bs * uni_stride + h]);
|
|
}
|
|
else
|
|
{
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] +=
|
|
((1 - activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + h],
|
|
2)) *
|
|
activfunc(
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h],
|
|
1));
|
|
}
|
|
}
|
|
else
|
|
{
|
|
|
|
pretime_shift = li * batch_n * hy_stride +
|
|
(bacc - in_n[ti - 1]) * hy_stride + bi * 3 * hy_h;
|
|
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] +=
|
|
((1 -
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + h], 2)) *
|
|
activfunc(
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h],
|
|
1) +
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + h], 2) *
|
|
hid_state[pretime_shift + bs * hy_stride + h]);
|
|
}
|
|
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + h], 2);
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + hy_h + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + hy_h + h], 2);
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
activfunc(hid_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h], 1);
|
|
|
|
hy_state[hx_shift + bs * uni_stride + h] =
|
|
hid_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h];
|
|
}
|
|
}
|
|
|
|
if(bidirection)
|
|
{
|
|
pretime_shift = li * batch_n * hy_stride +
|
|
(baccbi + in_n[seqLength - 1 - ti]) * hy_stride + bi * 3 * hy_h +
|
|
hy_h;
|
|
|
|
for(int bs = 0; bs < in_n[seqLength - 1 - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h +
|
|
h];
|
|
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + h] +=
|
|
activfunc(
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 4 * hy_h + h],
|
|
2) *
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h +
|
|
h];
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h +
|
|
h] = 0;
|
|
|
|
if(ti == 0)
|
|
{
|
|
if(!hx_is_null)
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h] +=
|
|
((1 -
|
|
activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2)) *
|
|
activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
5 * hy_h + h],
|
|
1) +
|
|
activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2) *
|
|
hx[hx_shift + bs * uni_stride + hy_n * hy_h + h]);
|
|
}
|
|
else
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h] +=
|
|
((1 -
|
|
activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2)) *
|
|
activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
5 * hy_h + h],
|
|
1));
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if(!hx_is_null && in_n.at(seqLength - 1 - ti) > in_n.at(seqLength - ti))
|
|
{
|
|
if(bs >= in_n.at(seqLength - ti))
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
bi * 3 * hy_h + hy_h + h] +=
|
|
(activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2) *
|
|
hx[hx_shift + bs * uni_stride + hy_n * hy_h + h]);
|
|
}
|
|
}
|
|
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h +
|
|
h] +=
|
|
((1 - activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2)) *
|
|
activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
5 * hy_h + h],
|
|
1));
|
|
|
|
if(bs < in_n[seqLength - ti])
|
|
{
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h] +=
|
|
(activfunc(hid_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2) *
|
|
hid_state[pretime_shift + bs * hy_stride + h]);
|
|
}
|
|
}
|
|
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
activfunc(
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h], 2);
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 4 * hy_h + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
activfunc(
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 4 * hy_h + h], 2);
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + h +
|
|
numlayer * batch_n * hy_stride] =
|
|
activfunc(
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + h], 1);
|
|
|
|
hy_state[hx_shift + bs * uni_stride + hy_n * hy_h + h] =
|
|
hid_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h +
|
|
h];
|
|
}
|
|
}
|
|
}
|
|
|
|
bacc += in_n[ti];
|
|
}
|
|
}
|
|
|
|
// output
|
|
int prelayer_shift = (numlayer - 1) * batch_n * hy_stride + bi * 3 * hy_h;
|
|
|
|
for(int i = 0; i < numlayer * batch_n * hy_stride * 2; i++)
|
|
{
|
|
rsvspace_host[i] = hid_state[i];
|
|
}
|
|
if(use_dropout)
|
|
{
|
|
for(int i = 0; i < (numlayer - 1) * batch_n * hy_h * bi; i++)
|
|
{
|
|
rsvspace_host.at(numlayer * batch_n * hy_stride * 2 + i) = dropout_hid_state.at(i);
|
|
}
|
|
auto p_drop_rsv =
|
|
reinterpret_cast<unsigned char*>(&rsvspace_host[numlayer * batch_n * hy_stride * 2 +
|
|
(numlayer - 1) * batch_n * hy_h * bi]);
|
|
for(int i = 0; i < (numlayer - 1) * batch_n * hy_h * bi; i++)
|
|
{
|
|
*(p_drop_rsv + i) = dropout_reservespace_host.at(i);
|
|
}
|
|
}
|
|
|
|
for(int i = 0; i < hy_d * hy_n * hy_h; i++)
|
|
{
|
|
hy_host[i] = hy_state[i];
|
|
}
|
|
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < out_h; h++)
|
|
{
|
|
out_host[bs * out_stride + h] = hid_state[prelayer_shift + bs * hy_stride + h];
|
|
}
|
|
}
|
|
|
|
delete[] out_state;
|
|
delete[] in_state;
|
|
delete[] hx_state;
|
|
delete[] hy_state;
|
|
delete[] wei_state;
|
|
}
|
|
|
|
template <typename Tgpu, typename Tref>
|
|
void RunGRUBackwardDataGEMMCPUVerify(std::vector<Tref>& din_host,
|
|
std::vector<Tgpu>& wei, // [ input_state_weight_trans
|
|
// hidden_state_weight0_trans input1_trans
|
|
// hidden1_trans ... output_weight;
|
|
// bidirectional reversed weights ]
|
|
std::vector<Tgpu>& dhy, // current/final hidden state
|
|
std::vector<Tref>& dhx_host,
|
|
std::vector<Tgpu>& hx, // initial hidden state
|
|
std::vector<Tgpu>& dout,
|
|
std::vector<int>& in_n, // input batch size
|
|
int in_h, // input data length
|
|
int seqLength, // Number of iterations to unroll over
|
|
bool bidirection, // whether using bidirectional net
|
|
bool biased, // whether using bias
|
|
int hy_d, // 1 by numlayer (number of stacks of hidden layers)
|
|
// for unidirection, 2 by numlayer for bidirection
|
|
int hy_n, // equal to input batch size in_n[0]
|
|
int hy_h, // hidden state number
|
|
int out_h, // 1 by hy_h related function for unidirection, 2 by
|
|
// hy_h related function for bidirection
|
|
int inputMode,
|
|
std::vector<Tref>& rsvspace_host,
|
|
std::vector<Tref>& wkspace_host,
|
|
bool use_dropout,
|
|
miopenDropoutDescriptor_t dropoutDesc,
|
|
bool hx_is_null = false,
|
|
bool dhy_is_null = false)
|
|
{
|
|
int batch_n = sumvc(in_n);
|
|
|
|
int numlayer = bidirection ? hy_d / 2 : hy_d;
|
|
int bacc, baccbi; // accumulation of batch
|
|
int bi = bidirection ? 2 : 1;
|
|
|
|
int in_stride = in_h;
|
|
int out_stride = out_h;
|
|
int wei_stride = bi * 3 * hy_h;
|
|
int hy_stride = bi * 4 * hy_h;
|
|
int h_stride = bi * hy_h;
|
|
int uni_stride = hy_h;
|
|
int bi_stride = hy_h * bi;
|
|
|
|
std::vector<Tref> dh_state(numlayer * batch_n * hy_stride, static_cast<Tref>(0));
|
|
|
|
Tref* din_state = new Tref[batch_n * in_h];
|
|
memset(din_state, 0, batch_n * in_h * sizeof(Tref));
|
|
|
|
// initial dout
|
|
Tref* dout_state = new Tref[batch_n * out_h];
|
|
for(int h = 0; h < batch_n; h++)
|
|
{
|
|
for(int w = 0; w < out_h; w++)
|
|
{
|
|
dout_state[h * out_stride + w] = dout[h * out_stride + w];
|
|
}
|
|
}
|
|
|
|
// initial hidden states
|
|
Tref* dhx_state = new Tref[hy_d * hy_n * hy_h];
|
|
memset(dhx_state, 0, hy_d * hy_n * hy_h * sizeof(Tref));
|
|
Tref* dcx_state = new Tref[hy_d * hy_n * hy_h];
|
|
memset(dcx_state, 0, hy_d * hy_n * hy_h * sizeof(Tref));
|
|
Tref* dhy_state = new Tref[hy_d * hy_n * hy_h];
|
|
for(int h = 0; h < hy_d * hy_n * hy_h; h++)
|
|
{
|
|
dhy_state[h] = dhy[h];
|
|
}
|
|
Tref* hx_state = new Tref[hy_d * hy_n * hy_h];
|
|
for(int h = 0; h < hy_d * hy_n * hy_h; h++)
|
|
{
|
|
hx_state[h] = hx[h];
|
|
}
|
|
|
|
if(inputMode == 1)
|
|
{
|
|
if(in_h != hy_h)
|
|
{
|
|
printf("Verification cannot be completed: The input tensor size must equal to the "
|
|
"hidden state size of the network in SKIP_INPUT mode!\n");
|
|
return;
|
|
}
|
|
in_h = 0;
|
|
}
|
|
|
|
int wei_len = (in_h + hy_h + (bi * hy_h + hy_h) * (numlayer - 1)) * wei_stride;
|
|
if(biased)
|
|
{
|
|
int in_bias = 2;
|
|
wei_len += (in_bias + (numlayer - 1) * 2) * wei_stride;
|
|
}
|
|
|
|
// initial weights
|
|
Tref* wei_state = new Tref[wei_len];
|
|
for(int h = 0; h < wei_len; h++)
|
|
{
|
|
wei_state[h] = wei[h];
|
|
}
|
|
|
|
// initial dropoput
|
|
miopenTensorDescriptor_t dropout_inputTensor{};
|
|
std::vector<unsigned char> dropout_reservespace_host;
|
|
if(use_dropout)
|
|
{
|
|
std::array<int, 2> drop_in_len = {{batch_n, hy_h * bi}};
|
|
std::array<int, 2> drop_in_str = {{hy_stride, 1}};
|
|
miopenCreateTensorDescriptor(&dropout_inputTensor);
|
|
miopenSetTensorDescriptor(
|
|
dropout_inputTensor, miopenFloat, 2, drop_in_len.data(), drop_in_str.data());
|
|
|
|
size_t reserveSpaceSizeInBytes = 0;
|
|
miopenDropoutGetReserveSpaceSize(dropout_inputTensor, &reserveSpaceSizeInBytes);
|
|
size_t reserve_size = reserveSpaceSizeInBytes / sizeof(unsigned char);
|
|
dropout_reservespace_host = std::vector<unsigned char>(reserve_size * (numlayer - 1),
|
|
static_cast<unsigned char>(0));
|
|
|
|
auto p_drop_rsv =
|
|
reinterpret_cast<unsigned char*>(&rsvspace_host[numlayer * batch_n * hy_stride * 2 +
|
|
(numlayer - 1) * batch_n * hy_h * bi]);
|
|
for(int i = 0; i < (numlayer - 1) * batch_n * hy_h * bi; i++)
|
|
{
|
|
dropout_reservespace_host.at(i) = *(p_drop_rsv + i);
|
|
}
|
|
}
|
|
|
|
// bwd data emulator
|
|
for(int li = numlayer - 1; li >= 0; li--)
|
|
{
|
|
int wei_shift = (in_h + hy_h) * wei_stride + li * (bi * hy_h + hy_h) * wei_stride;
|
|
int hid_shift = li * batch_n * hy_stride;
|
|
int hx_shift = li * in_n[0] * h_stride;
|
|
int weitime_shift = in_h * wei_stride + li * (bi * hy_h + hy_h) * wei_stride;
|
|
|
|
if(li == numlayer - 1)
|
|
{
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < out_h; h++)
|
|
{
|
|
dh_state[hid_shift + bi * 3 * hy_h + bs * hy_stride + h] +=
|
|
dout_state[bs * out_stride + h];
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
int prelayer_shift = (li + 1) * batch_n * hy_stride;
|
|
|
|
gemm_cpu<Tref>(&dh_state[prelayer_shift],
|
|
hy_h * bi * 3,
|
|
batch_n,
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[wei_shift]),
|
|
hy_h * bi,
|
|
hy_h * bi * 3,
|
|
bi_stride,
|
|
false,
|
|
&dh_state[hid_shift + bi * 3 * hy_h],
|
|
hy_h * bi,
|
|
batch_n,
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
if(use_dropout)
|
|
{
|
|
RunDropoutBackwardEmulator<Tref>(dropoutDesc,
|
|
dropout_inputTensor,
|
|
dh_state,
|
|
dropout_inputTensor,
|
|
dh_state,
|
|
dropout_reservespace_host,
|
|
hid_shift + bi * 3 * hy_h,
|
|
hid_shift + bi * 3 * hy_h,
|
|
li * batch_n * hy_h * bi);
|
|
}
|
|
}
|
|
|
|
// from hidden state
|
|
bacc = batch_n;
|
|
baccbi = 0;
|
|
for(int ti = seqLength - 1; ti >= 0; ti--)
|
|
{
|
|
bacc -= in_n[ti];
|
|
|
|
if(ti == seqLength - 1)
|
|
{
|
|
if(!dhy_is_null)
|
|
{
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] +=
|
|
dhy_state[hx_shift + bs * uni_stride + h];
|
|
}
|
|
}
|
|
|
|
if(bidirection)
|
|
{
|
|
for(int bs = 0; bs < in_n[seqLength - 1 - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h] +=
|
|
dhy_state[hx_shift + bs * uni_stride + hy_n * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if(!dhy_is_null && in_n.at(ti) > in_n.at(ti + 1))
|
|
{
|
|
for(int bs = in_n.at(ti + 1); bs < in_n.at(ti); bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] +=
|
|
dhy_state[hx_shift + bs * uni_stride + h];
|
|
}
|
|
}
|
|
}
|
|
|
|
int pretime_shift = li * batch_n * hy_stride + (bacc + in_n[ti]) * hy_stride;
|
|
|
|
gemm_cpu<Tref>(&dh_state[pretime_shift],
|
|
hy_h * 2,
|
|
in_n[ti + 1],
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
false,
|
|
&dh_state[hid_shift + bacc * hy_stride + bi * 3 * hy_h],
|
|
hy_h,
|
|
in_n[ti + 1],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
for(int bs = 0; bs < in_n[ti + 1]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] +=
|
|
dh_state[pretime_shift + bs * hy_stride + bi * 3 * hy_h + h] *
|
|
activfunc(rsvspace_host[pretime_shift + bs * hy_stride + h], 2);
|
|
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h] =
|
|
dh_state[pretime_shift + bs * hy_stride + 2 * hy_h + h] *
|
|
activfunc(rsvspace_host[pretime_shift + bs * hy_stride + hy_h + h], 2);
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(&dh_state[hid_shift + bacc * hy_stride + 2 * hy_h],
|
|
hy_h,
|
|
in_n[ti + 1],
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift + 2 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
false,
|
|
&dh_state[hid_shift + bacc * hy_stride + bi * 3 * hy_h],
|
|
hy_h,
|
|
in_n[ti + 1],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
for(int bs = 0; bs < in_n[ti + 1]; bs++)
|
|
{
|
|
for(int hs = 0; hs < hy_h; hs++)
|
|
{
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + hs] = 0;
|
|
}
|
|
}
|
|
|
|
if(bidirection)
|
|
{
|
|
pretime_shift = li * batch_n * hy_stride +
|
|
(baccbi - in_n[seqLength - 2 - ti]) * hy_stride + hy_h * 3;
|
|
|
|
gemm_cpu<Tref>(
|
|
&dh_state[pretime_shift],
|
|
hy_h * 2,
|
|
in_n[seqLength - 1 - ti],
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift + hy_h * 3 * uni_stride]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
false,
|
|
&dh_state[hid_shift + baccbi * hy_stride + bi * 3 * hy_h + hy_h],
|
|
hy_h,
|
|
in_n[seqLength - 1 - ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
for(int bs = 0; bs < in_n[seqLength - 1 - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h +
|
|
h] +=
|
|
dh_state[pretime_shift + bs * hy_stride + 3 * hy_h + hy_h + h] *
|
|
activfunc(rsvspace_host[pretime_shift + bs * hy_stride + h], 2);
|
|
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + h] =
|
|
dh_state[pretime_shift + bs * hy_stride + 2 * hy_h + h] *
|
|
activfunc(rsvspace_host[pretime_shift + bs * hy_stride + hy_h + h],
|
|
2);
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(
|
|
&dh_state[hid_shift + baccbi * hy_stride + 5 * hy_h],
|
|
hy_h,
|
|
in_n[seqLength - 1 - ti],
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift + 5 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
false,
|
|
&dh_state[hid_shift + baccbi * hy_stride + bi * 3 * hy_h + hy_h],
|
|
hy_h,
|
|
in_n[seqLength - 1 - ti],
|
|
hy_stride,
|
|
1,
|
|
1);
|
|
|
|
for(int bs = 0; bs < in_n[seqLength - 1 - ti]; bs++)
|
|
{
|
|
for(int hs = 0; hs < hy_h; hs++)
|
|
{
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + hs] = 0;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h] +=
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] *
|
|
(1 - activfunc(rsvspace_host[hid_shift + (bacc + bs) * hy_stride + h], 2)) *
|
|
dervactivfunc(
|
|
rsvspace_host[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h], 1);
|
|
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + hy_h + h] =
|
|
(rsvspace_host[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h +
|
|
numlayer * batch_n * hy_stride] *
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h] *
|
|
dervactivfunc(
|
|
rsvspace_host[hid_shift + (bacc + bs) * hy_stride + hy_h + h], 2));
|
|
|
|
if(ti == 0)
|
|
{
|
|
if(!hx_is_null)
|
|
{
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + h] +=
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] *
|
|
hx_state[hx_shift + bs * uni_stride + h] *
|
|
dervactivfunc(
|
|
rsvspace_host[hid_shift + (bacc + bs) * hy_stride + h], 2);
|
|
}
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + h] -=
|
|
(dh_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] *
|
|
activfunc(
|
|
rsvspace_host[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h],
|
|
1) *
|
|
dervactivfunc(rsvspace_host[hid_shift + (bacc + bs) * hy_stride + h],
|
|
2));
|
|
}
|
|
else
|
|
{
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + h] +=
|
|
dh_state[hid_shift + (bacc + bs) * hy_stride + bi * 3 * hy_h + h] *
|
|
(rsvspace_host[hid_shift + (bacc - in_n[ti - 1] + bs) * hy_stride +
|
|
bi * 3 * hy_h + h] -
|
|
activfunc(
|
|
rsvspace_host[hid_shift + (bacc + bs) * hy_stride + 2 * hy_h + h],
|
|
1)) *
|
|
dervactivfunc(rsvspace_host[hid_shift + (bacc + bs) * hy_stride + h],
|
|
2);
|
|
}
|
|
}
|
|
}
|
|
|
|
if(bidirection)
|
|
{
|
|
for(int bs = 0; bs < in_n[seqLength - 1 - ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + h] +=
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h + hy_h +
|
|
h] *
|
|
(1 - activfunc(rsvspace_host[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2)) *
|
|
dervactivfunc(
|
|
rsvspace_host[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + h],
|
|
1);
|
|
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 4 * hy_h + h] =
|
|
rsvspace_host[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h + numlayer * batch_n * hy_stride];
|
|
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 4 * hy_h + h] *=
|
|
(dh_state[hid_shift + (baccbi + bs) * hy_stride + 5 * hy_h + h] *
|
|
dervactivfunc(rsvspace_host[hid_shift + (baccbi + bs) * hy_stride +
|
|
4 * hy_h + h],
|
|
2));
|
|
|
|
if(ti == 0)
|
|
{
|
|
if(!hx_is_null)
|
|
{
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h] +=
|
|
(dh_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
bi * 3 * hy_h + hy_h + h] *
|
|
hx_state[hx_shift + bs * uni_stride + hy_n * hy_h + h]);
|
|
}
|
|
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h] -=
|
|
(dh_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h] *
|
|
activfunc(rsvspace_host[hid_shift + (baccbi + bs) * hy_stride +
|
|
5 * hy_h + h],
|
|
1));
|
|
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h] *=
|
|
dervactivfunc(rsvspace_host[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2);
|
|
}
|
|
else
|
|
{
|
|
if(!hx_is_null &&
|
|
in_n.at(seqLength - 1 - ti) > in_n.at(seqLength - ti) &&
|
|
bs >= in_n.at(seqLength - ti))
|
|
{
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h] +=
|
|
(dh_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
bi * 3 * hy_h + hy_h + h] *
|
|
hx_state[hx_shift + bs * uni_stride + hy_n * hy_h + h]);
|
|
}
|
|
|
|
if(bs < in_n[seqLength - ti])
|
|
{
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h] +=
|
|
(dh_state[hid_shift + (baccbi + bs) * hy_stride +
|
|
bi * 3 * hy_h + hy_h + h] *
|
|
rsvspace_host[hid_shift +
|
|
(baccbi + in_n[seqLength - 1 - ti] + bs) *
|
|
hy_stride +
|
|
bi * 3 * hy_h + hy_h + h]);
|
|
}
|
|
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h] -=
|
|
(dh_state[hid_shift + (baccbi + bs) * hy_stride + bi * 3 * hy_h +
|
|
hy_h + h] *
|
|
activfunc(rsvspace_host[hid_shift + (baccbi + bs) * hy_stride +
|
|
5 * hy_h + h],
|
|
1));
|
|
|
|
dh_state[hid_shift + (baccbi + bs) * hy_stride + 3 * hy_h + h] *=
|
|
dervactivfunc(rsvspace_host[hid_shift + (baccbi + bs) * hy_stride +
|
|
3 * hy_h + h],
|
|
2);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
baccbi += in_n[seqLength - 1 - ti];
|
|
}
|
|
|
|
// dhx
|
|
int pretime_shift = li * batch_n * hy_stride;
|
|
|
|
gemm_cpu<Tref>(&dh_state[pretime_shift],
|
|
hy_h * 2,
|
|
in_n[0],
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
false,
|
|
&dhx_state[hx_shift],
|
|
hy_h,
|
|
in_n[0],
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
for(int bs = 0; bs < in_n[0]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dhx_state[hx_shift + bs * uni_stride + h] +=
|
|
dh_state[pretime_shift + bs * hy_stride + bi * 3 * hy_h + h] *
|
|
activfunc(rsvspace_host[pretime_shift + bs * hy_stride + h], 2);
|
|
|
|
dcx_state[hx_shift + bs * uni_stride + h] =
|
|
dh_state[pretime_shift + bs * hy_stride + 2 * hy_h + h] *
|
|
activfunc(rsvspace_host[pretime_shift + bs * hy_stride + hy_h + h], 2);
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&dcx_state[hx_shift]),
|
|
hy_h,
|
|
in_n[0],
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift + 2 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
false,
|
|
&dhx_state[hx_shift],
|
|
hy_h,
|
|
in_n[0],
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
if(bidirection)
|
|
{
|
|
int ti = seqLength - 1, cur_bat = 0, pre_bat = batch_n;
|
|
|
|
while(ti >= 0)
|
|
{
|
|
pre_bat -= in_n.at(ti);
|
|
if(in_n.at(ti) > cur_bat)
|
|
{
|
|
pretime_shift = li * batch_n * hy_stride + (pre_bat + cur_bat) * hy_stride;
|
|
|
|
gemm_cpu<Tref>(
|
|
&dh_state[pretime_shift + 3 * hy_h],
|
|
hy_h * 2,
|
|
(in_n.at(ti) - cur_bat),
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift + 3 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h * 2,
|
|
uni_stride,
|
|
false,
|
|
&dhx_state[hx_shift + hy_n * hy_h + cur_bat * hy_h],
|
|
hy_h,
|
|
(in_n.at(ti) - cur_bat),
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
for(int bs = cur_bat; bs < in_n.at(ti); bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
dhx_state[hx_shift + bs * uni_stride + hy_n * hy_h + h] +=
|
|
dh_state[pretime_shift + (bs - cur_bat) * hy_stride +
|
|
bi * 3 * hy_h + hy_h + h] *
|
|
activfunc(rsvspace_host[pretime_shift + (bs - cur_bat) * hy_stride +
|
|
3 * hy_h + h],
|
|
2);
|
|
|
|
dcx_state[hx_shift + bs * uni_stride + hy_n * hy_h + h] =
|
|
dh_state[pretime_shift + (bs - cur_bat) * hy_stride + 5 * hy_h +
|
|
h] *
|
|
activfunc(rsvspace_host[pretime_shift + (bs - cur_bat) * hy_stride +
|
|
4 * hy_h + h],
|
|
2);
|
|
}
|
|
}
|
|
|
|
gemm_cpu<Tref>(
|
|
const_cast<Tref*>(&dcx_state[hx_shift + hy_n * hy_h + cur_bat * hy_h]),
|
|
hy_h,
|
|
(in_n.at(ti) - cur_bat),
|
|
uni_stride,
|
|
false,
|
|
const_cast<Tref*>(&wei_state[weitime_shift + 5 * hy_h * uni_stride]),
|
|
hy_h,
|
|
hy_h,
|
|
uni_stride,
|
|
false,
|
|
&dhx_state[hx_shift + hy_n * hy_h + cur_bat * hy_h],
|
|
hy_h,
|
|
(in_n.at(ti) - cur_bat),
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
}
|
|
cur_bat = in_n.at(ti--);
|
|
}
|
|
}
|
|
}
|
|
|
|
// dinput
|
|
if(inputMode == 1)
|
|
{
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
for(int gi = 0; gi < 3; gi++)
|
|
{
|
|
din_state[bs * in_stride + h] += dh_state[bs * hy_stride + gi * hy_h + h];
|
|
if(bidirection)
|
|
{
|
|
din_state[bs * in_stride + h] +=
|
|
dh_state[bs * hy_stride + (gi + 3) * hy_h + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
gemm_cpu<Tref>(&dh_state[0],
|
|
hy_h * bi * 3,
|
|
batch_n,
|
|
hy_stride,
|
|
false,
|
|
const_cast<Tref*>(wei_state),
|
|
in_h,
|
|
hy_h * bi * 3,
|
|
in_stride,
|
|
false,
|
|
&din_state[0],
|
|
in_h,
|
|
batch_n,
|
|
in_stride,
|
|
1,
|
|
1);
|
|
}
|
|
|
|
for(int i = 0; i < numlayer * batch_n * hy_stride; i++)
|
|
{
|
|
wkspace_host[i] = dh_state[i];
|
|
}
|
|
|
|
for(int i = 0; i < hy_d * hy_n * hy_h; i++)
|
|
{
|
|
dhx_host[i] = dhx_state[i];
|
|
}
|
|
|
|
for(int bs = 0; bs < batch_n; bs++)
|
|
{
|
|
for(int h = 0; h < in_stride; h++)
|
|
{
|
|
din_host[bs * in_stride + h] = din_state[bs * in_stride + h];
|
|
}
|
|
}
|
|
|
|
delete[] dout_state;
|
|
delete[] din_state;
|
|
delete[] hx_state;
|
|
delete[] dhx_state;
|
|
delete[] dcx_state;
|
|
delete[] dhy_state;
|
|
delete[] wei_state;
|
|
}
|
|
|
|
template <typename Tgpu, typename Tref>
|
|
void RunGRUBackwardWeightGEMMCPUVerify(std::vector<Tgpu>& in,
|
|
std::vector<Tref>& dwei_host, // [ input_state_weight_trans
|
|
// hidden_state_weight0_trans
|
|
// input1_trans hidden1_trans ...
|
|
// output_weight; bidirectional
|
|
// reversed weights ]
|
|
std::vector<Tgpu>& hx, // initial hidden state
|
|
std::vector<Tgpu>& dout,
|
|
std::vector<int>& in_n, // input batch size
|
|
int in_h, // input data length
|
|
int seqLength, // Number of iterations to unroll over
|
|
bool bidirection, // whether using bidirectional net
|
|
bool biased, // whether using bias
|
|
int hy_d, // 1 by numlayer (number of stacks of hidden
|
|
// layers) for unidirection, 2 by numlayer for
|
|
// bidirection
|
|
int hy_n, // equal to input batch size in_n[0]
|
|
int hy_h, // hidden state number
|
|
int out_h, // 1 by hy_h related function for unidirection, 2
|
|
// by hy_h related function for bidirection
|
|
int inputMode,
|
|
std::vector<Tref>& rsvspace_host,
|
|
std::vector<Tref>& wkspace_host,
|
|
bool use_dropout,
|
|
bool hx_is_null = false)
|
|
{
|
|
int batch_n = sumvc(in_n);
|
|
int numlayer = bidirection ? hy_d / 2 : hy_d;
|
|
int bacc; // accumulation of batch
|
|
int bi = bidirection ? 2 : 1;
|
|
|
|
int in_stride = in_h;
|
|
int wei_stride = bi * 3 * hy_h;
|
|
int hy_stride = bi * 4 * hy_h;
|
|
int h_stride = bi * hy_h;
|
|
int uni_stride = hy_h;
|
|
int bi_stride = hy_h * bi;
|
|
|
|
// initial input
|
|
Tref* in_state = new Tref[batch_n * in_h];
|
|
for(int h = 0; h < batch_n; h++)
|
|
{
|
|
for(int w = 0; w < in_h; w++)
|
|
{
|
|
in_state[h * in_h + w] = in[h * in_h + w];
|
|
}
|
|
}
|
|
|
|
// initial output difference
|
|
Tref* dout_state = new Tref[batch_n * out_h];
|
|
for(int h = 0; h < batch_n; h++)
|
|
{
|
|
for(int w = 0; w < out_h; w++)
|
|
{
|
|
dout_state[h * out_h + w] = dout[h * out_h + w];
|
|
}
|
|
}
|
|
|
|
// initial saved data
|
|
Tref* wkspace_state = new Tref[numlayer * batch_n * hy_stride];
|
|
for(int h = 0; h < numlayer * batch_n * hy_stride; h++)
|
|
{
|
|
wkspace_state[h] = wkspace_host[h];
|
|
}
|
|
size_t rsvsp_sz = use_dropout ? rsvspace_host.size() : numlayer * batch_n * hy_stride;
|
|
Tref* rsvspace_state = new Tref[rsvsp_sz];
|
|
for(int h = 0; h < rsvsp_sz; h++)
|
|
{
|
|
rsvspace_state[h] = rsvspace_host[h];
|
|
}
|
|
|
|
// initial hidden states
|
|
Tref* hx_state = new Tref[hy_d * hy_n * hy_h];
|
|
for(int h = 0; h < hy_d * hy_n * hy_h; h++)
|
|
{
|
|
hx_state[h] = hx[h];
|
|
}
|
|
|
|
if(inputMode == 1)
|
|
{
|
|
if(in_h != hy_h)
|
|
{
|
|
printf("Verification cannot be completed: The input tensor size must equal to the "
|
|
"hidden state size of the network in SKIP_INPUT mode!\n");
|
|
return;
|
|
}
|
|
in_h = 0;
|
|
}
|
|
|
|
int wei_shift_bias = (in_h + hy_h + (bi * hy_h + hy_h) * (numlayer - 1)) * wei_stride;
|
|
int wei_len = wei_shift_bias;
|
|
if(biased)
|
|
{
|
|
int in_bias = 2;
|
|
wei_len += (in_bias + (numlayer - 1) * 2) * wei_stride;
|
|
}
|
|
|
|
// initial dwei
|
|
Tref* dwei_state = new Tref[wei_len];
|
|
memset(dwei_state, 0, wei_len * sizeof(Tref));
|
|
|
|
// bwd weights emulator
|
|
for(int li = 0; li < numlayer; li++)
|
|
{
|
|
// between layers
|
|
if(li == 0)
|
|
{
|
|
if(inputMode == 0)
|
|
{
|
|
gemm_cpu<Tref>(const_cast<Tref*>(wkspace_state),
|
|
hy_h * bi * 3,
|
|
batch_n,
|
|
hy_stride,
|
|
true,
|
|
const_cast<Tref*>(in_state),
|
|
in_h,
|
|
batch_n,
|
|
in_stride,
|
|
false,
|
|
&dwei_state[0],
|
|
in_h,
|
|
hy_h * bi * 3,
|
|
in_stride,
|
|
1,
|
|
1);
|
|
}
|
|
|
|
if(biased)
|
|
{
|
|
for(int h = 0; h < wei_stride; h++)
|
|
{
|
|
for(int w = 0; w < batch_n; w++)
|
|
{
|
|
dwei_state[wei_shift_bias + h] += wkspace_state[w * hy_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
int prelayer_shift =
|
|
use_dropout ? 2 * numlayer * batch_n * hy_stride + (li - 1) * batch_n * hy_h * bi
|
|
: (li - 1) * batch_n * hy_stride + bi * hy_h * 3;
|
|
int hid_shift = li * batch_n * hy_stride;
|
|
int wei_shift = (in_h + hy_h) * wei_stride + (li - 1) * (bi * hy_h + hy_h) * wei_stride;
|
|
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&wkspace_state[hid_shift]),
|
|
hy_h * bi * 3,
|
|
batch_n,
|
|
hy_stride,
|
|
true,
|
|
const_cast<Tref*>(&rsvspace_state[prelayer_shift]),
|
|
hy_h * bi,
|
|
batch_n,
|
|
use_dropout ? hy_h * bi : hy_stride,
|
|
false,
|
|
&dwei_state[wei_shift],
|
|
hy_h * bi,
|
|
hy_h * bi * 3,
|
|
bi_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
wei_shift = wei_shift_bias + li * 2 * wei_stride;
|
|
|
|
for(int h = 0; h < wei_stride; h++)
|
|
{
|
|
for(int w = 0; w < batch_n; w++)
|
|
{
|
|
dwei_state[wei_shift + h] += wkspace_state[hid_shift + w * hy_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
// between time
|
|
bacc = 0;
|
|
for(int ti = 0; ti < seqLength; ti++)
|
|
{
|
|
int hid_shift = li * batch_n * hy_stride + bacc * hy_stride;
|
|
int hx_shift = li * in_n[0] * h_stride;
|
|
int wei_shift = in_h * wei_stride + li * (bi * hy_h + hy_h) * wei_stride;
|
|
int pretime_shift;
|
|
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
wkspace_state[hid_shift + bs * hy_stride + 2 * hy_h + h] *=
|
|
activfunc(rsvspace_state[hid_shift + bs * hy_stride + hy_h + h], 2);
|
|
}
|
|
}
|
|
|
|
// between time
|
|
if(ti == 0)
|
|
{
|
|
if(!hx_is_null)
|
|
{
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&wkspace_state[hid_shift]),
|
|
hy_h * 3,
|
|
in_n[ti],
|
|
hy_stride,
|
|
true,
|
|
const_cast<Tref*>(&hx_state[hx_shift]),
|
|
hy_h,
|
|
in_n[ti],
|
|
uni_stride,
|
|
false,
|
|
&dwei_state[wei_shift],
|
|
hy_h,
|
|
hy_h * 3,
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
int bias_shift = wei_shift_bias + li * 2 * wei_stride + wei_stride;
|
|
|
|
for(int h = 0; h < hy_h * 3; h++)
|
|
{
|
|
for(int w = 0; w < in_n.at(ti); w++)
|
|
{
|
|
dwei_state[bias_shift + h] +=
|
|
wkspace_state[hid_shift + w * hy_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
pretime_shift =
|
|
li * batch_n * hy_stride + (bacc - in_n[ti - 1]) * hy_stride + bi * 3 * hy_h;
|
|
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&wkspace_state[hid_shift]),
|
|
hy_h * 3,
|
|
in_n[ti],
|
|
hy_stride,
|
|
true,
|
|
const_cast<Tref*>(&rsvspace_state[pretime_shift]),
|
|
hy_h,
|
|
in_n[ti],
|
|
hy_stride,
|
|
false,
|
|
&dwei_state[wei_shift],
|
|
hy_h,
|
|
hy_h * 3,
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
int bias_shift = wei_shift_bias + li * 2 * wei_stride + wei_stride;
|
|
|
|
for(int h = 0; h < hy_h * 3; h++)
|
|
{
|
|
for(int w = 0; w < in_n.at(ti); w++)
|
|
{
|
|
dwei_state[bias_shift + h] +=
|
|
wkspace_state[hid_shift + w * hy_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
if(bidirection)
|
|
{
|
|
for(int bs = 0; bs < in_n[ti]; bs++)
|
|
{
|
|
for(int h = 0; h < hy_h; h++)
|
|
{
|
|
wkspace_state[hid_shift + bs * hy_stride + 5 * hy_h + h] *=
|
|
activfunc(rsvspace_state[hid_shift + bs * hy_stride + 4 * hy_h + h], 2);
|
|
}
|
|
}
|
|
|
|
if(ti == seqLength - 1)
|
|
{
|
|
if(!hx_is_null)
|
|
{
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&wkspace_state[hid_shift + 3 * hy_h]),
|
|
hy_h * 3,
|
|
in_n[ti],
|
|
hy_stride,
|
|
true,
|
|
const_cast<Tref*>(&hx_state[hx_shift + hy_n * hy_h]),
|
|
hy_h,
|
|
in_n[ti],
|
|
uni_stride,
|
|
false,
|
|
&dwei_state[wei_shift + 3 * hy_h * uni_stride],
|
|
hy_h,
|
|
hy_h * 3,
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
int bias_shift = wei_shift_bias + li * 2 * wei_stride + wei_stride;
|
|
|
|
for(int h = 0; h < hy_h * 3; h++)
|
|
{
|
|
for(int w = 0; w < in_n.at(ti); w++)
|
|
{
|
|
dwei_state[bias_shift + 3 * hy_h + h] +=
|
|
wkspace_state[hid_shift + 3 * hy_h + w * hy_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if(!hx_is_null && in_n.at(ti) > in_n.at(ti + 1))
|
|
{
|
|
gemm_cpu<Tref>(
|
|
const_cast<Tref*>(
|
|
&wkspace_state[hid_shift + 3 * hy_h + in_n.at(ti + 1) * hy_stride]),
|
|
hy_h * 3,
|
|
(in_n.at(ti) - in_n.at(ti + 1)),
|
|
hy_stride,
|
|
true,
|
|
const_cast<Tref*>(
|
|
&hx_state[hx_shift + hy_n * hy_h + in_n.at(ti + 1) * hy_h]),
|
|
hy_h,
|
|
(in_n.at(ti) - in_n.at(ti + 1)),
|
|
uni_stride,
|
|
false,
|
|
&dwei_state[wei_shift + 3 * hy_h * uni_stride],
|
|
hy_h,
|
|
hy_h * 3,
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
int bias_shift = wei_shift_bias + li * 2 * wei_stride + wei_stride;
|
|
|
|
for(int h = 0; h < hy_h * 3; h++)
|
|
{
|
|
for(int w = in_n.at(ti + 1); w < in_n.at(ti); w++)
|
|
{
|
|
dwei_state[bias_shift + 3 * hy_h + h] +=
|
|
wkspace_state[hid_shift + 3 * hy_h + w * hy_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
pretime_shift =
|
|
li * batch_n * hy_stride + (bacc + in_n[ti]) * hy_stride + bi * 3 * hy_h;
|
|
|
|
gemm_cpu<Tref>(const_cast<Tref*>(&wkspace_state[hid_shift + 3 * hy_h]),
|
|
hy_h * 3,
|
|
in_n[ti + 1],
|
|
hy_stride,
|
|
true,
|
|
const_cast<Tref*>(&rsvspace_state[pretime_shift + hy_h]),
|
|
hy_h,
|
|
in_n[ti + 1],
|
|
hy_stride,
|
|
false,
|
|
&dwei_state[wei_shift + 3 * hy_h * uni_stride],
|
|
hy_h,
|
|
hy_h * 3,
|
|
uni_stride,
|
|
1,
|
|
1);
|
|
|
|
if(biased)
|
|
{
|
|
int bias_shift = wei_shift_bias + li * 2 * wei_stride + wei_stride;
|
|
|
|
for(int h = 0; h < hy_h * 3; h++)
|
|
{
|
|
for(int w = 0; w < in_n.at(ti + 1); w++)
|
|
{
|
|
dwei_state[bias_shift + 3 * hy_h + h] +=
|
|
wkspace_state[hid_shift + 3 * hy_h + w * hy_stride + h];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
bacc += in_n[ti];
|
|
}
|
|
}
|
|
|
|
for(int i = 0; i < wei_len; i++)
|
|
{
|
|
dwei_host[i] = dwei_state[i];
|
|
}
|
|
|
|
delete[] dwei_state;
|
|
delete[] in_state;
|
|
delete[] dout_state;
|
|
delete[] wkspace_state;
|
|
delete[] rsvspace_state;
|
|
delete[] hx_state;
|
|
}
|
|
|
|
#endif // GUARD_MIOPEN_GRU_VERIFY_GEMM_HPP
|