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
587 lines
24 KiB
C++
587 lines
24 KiB
C++
/*******************************************************************************
|
|
*
|
|
* MIT License
|
|
*
|
|
* Copyright (c) 2019 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.
|
|
*
|
|
*******************************************************************************/
|
|
#pragma once
|
|
|
|
#include <cmath>
|
|
#include <cassert>
|
|
#include <algorithm>
|
|
#include <cstdlib>
|
|
#include <cstring>
|
|
#include <cfloat>
|
|
#include <vector>
|
|
#include "ctc_gpu_emulator.hpp"
|
|
|
|
#define NEGATIVE_CUTOFF_VAL (-1e20)
|
|
|
|
template <typename T>
|
|
T logaddexp(T x, T y)
|
|
{
|
|
T a = std::max(x, y);
|
|
T b = std::min(x, y);
|
|
T c = b - a;
|
|
|
|
return c <= T(NEGATIVE_CUTOFF_VAL)
|
|
? std::max(a, T(NEGATIVE_CUTOFF_VAL))
|
|
: std::max(T(a + log(T(1) + exp(b - a))), T(NEGATIVE_CUTOFF_VAL));
|
|
}
|
|
|
|
template <typename T>
|
|
T logsumexp(std::vector<T>& in_vec)
|
|
{
|
|
auto sum = in_vec[0];
|
|
for(int i = 1; i < in_vec.size(); i++)
|
|
sum = logaddexp(*(in_vec.begin() + i), sum);
|
|
|
|
return sum;
|
|
}
|
|
|
|
template <typename Tgpu, typename Tref = Tgpu>
|
|
void subvec_logsoftmax(std::vector<Tgpu>& in,
|
|
std::vector<Tref>& out,
|
|
size_t in_offset,
|
|
size_t out_offset,
|
|
size_t length)
|
|
{
|
|
auto itr_in = in.begin() + in_offset;
|
|
auto itr_out = out.begin() + out_offset;
|
|
|
|
std::vector<Tgpu> sub_in(itr_in, itr_in + length);
|
|
Tgpu max_val = *std::max_element(sub_in.begin(), sub_in.end());
|
|
for(int i = 0; i < length; i++)
|
|
sub_in[i] -= max_val;
|
|
|
|
Tgpu sum = logsumexp(sub_in);
|
|
for(int i = 0; i < length; i++)
|
|
*(itr_out + i) = std::max(Tref(sub_in[i] - sum), Tref(NEGATIVE_CUTOFF_VAL));
|
|
}
|
|
|
|
template <typename T>
|
|
std::vector<T> ctc_forward_log(std::vector<T>& probs_logits,
|
|
std::vector<int>& label,
|
|
const int input_length,
|
|
const int batch_size,
|
|
const int class_sz,
|
|
const int pstr0,
|
|
const int pstr1,
|
|
const int pstr2,
|
|
const int batch_id,
|
|
int blank_lb = 0)
|
|
{
|
|
(void)batch_size;
|
|
int probs_stride[] = {pstr0, pstr1, pstr2};
|
|
int label_prime_len = 2 * label.size() + 1;
|
|
blank_lb = blank_lb < 0 ? 0 : (blank_lb >= class_sz ? class_sz - 1 : blank_lb);
|
|
std::vector<int> label_prime(label_prime_len, blank_lb);
|
|
for(int i = 0; i < label.size(); i++)
|
|
label_prime[2 * i + 1] = label[i];
|
|
|
|
std::vector<T> alpha(input_length * label_prime_len, T(NEGATIVE_CUTOFF_VAL));
|
|
|
|
for(int j = 0; j < input_length; j++)
|
|
for(int i = 0; i < label_prime_len; i++)
|
|
{
|
|
size_t pidx = j * probs_stride[0] + batch_id * probs_stride[1] + label_prime[i];
|
|
size_t aidx_ts = j * label_prime_len + i;
|
|
if(j == 0 && i > 1)
|
|
break;
|
|
else if(j == 0)
|
|
{
|
|
alpha[aidx_ts] = probs_logits[pidx];
|
|
continue;
|
|
}
|
|
size_t aidx_t1s = (j - 1) * label_prime_len + i;
|
|
alpha[aidx_ts] = alpha[aidx_t1s];
|
|
if(i >= 1)
|
|
alpha[aidx_ts] = logaddexp(alpha[aidx_ts], alpha[aidx_t1s - 1]);
|
|
if(i >= 2)
|
|
if(label_prime[i] != blank_lb && label_prime[i] != label_prime[i - 2])
|
|
alpha[aidx_ts] = logaddexp(alpha[aidx_ts], alpha[aidx_t1s - 2]);
|
|
|
|
alpha[aidx_ts] += probs_logits[pidx];
|
|
alpha[aidx_ts] = std::max(alpha[aidx_ts], T(NEGATIVE_CUTOFF_VAL));
|
|
}
|
|
|
|
return alpha;
|
|
}
|
|
|
|
template <typename T>
|
|
std::vector<T> ctc_backward_log(std::vector<T>& probs_logits,
|
|
std::vector<int>& label,
|
|
const int input_length,
|
|
const int batch_size,
|
|
const int class_sz,
|
|
const int pstr0,
|
|
const int pstr1,
|
|
const int pstr2,
|
|
const int batch_id,
|
|
int blank_lb = 0)
|
|
{
|
|
(void)batch_size;
|
|
int probs_stride[] = {pstr0, pstr1, pstr2};
|
|
int label_prime_len = 2 * label.size() + 1;
|
|
blank_lb = blank_lb < 0 ? 0 : (blank_lb >= class_sz ? class_sz - 1 : blank_lb);
|
|
std::vector<int> label_prime(label_prime_len, blank_lb);
|
|
for(int i = 0; i < label.size(); i++)
|
|
label_prime[2 * i + 1] = label[i];
|
|
|
|
std::vector<T> beta(input_length * label_prime_len, T(NEGATIVE_CUTOFF_VAL));
|
|
|
|
for(int j = 0; j < input_length; j++)
|
|
for(int i = 0; i < label_prime_len; i++)
|
|
{
|
|
int j1 = input_length - 1 - j;
|
|
int i1 = label_prime_len - 1 - i;
|
|
size_t pidx = j1 * probs_stride[0] + batch_id * probs_stride[1] + label_prime[i1];
|
|
size_t bidx_ts = j1 * label_prime_len + i1;
|
|
if(j == 0 && i > 1)
|
|
break;
|
|
else if(j == 0)
|
|
{
|
|
beta[bidx_ts] = probs_logits[pidx];
|
|
continue;
|
|
}
|
|
size_t bidx_t1s = (j1 + 1) * label_prime_len + i1;
|
|
|
|
beta[bidx_ts] = beta[bidx_t1s];
|
|
if(i1 <= label_prime_len - 2)
|
|
beta[bidx_ts] = logaddexp(beta[bidx_ts], beta[bidx_t1s + 1]);
|
|
if(i1 <= label_prime_len - 3)
|
|
if(label_prime[i1] != blank_lb && label_prime[i1] != label_prime[i1 + 2])
|
|
beta[bidx_ts] = logaddexp(beta[bidx_ts], beta[bidx_t1s + 2]);
|
|
|
|
beta[bidx_ts] += probs_logits[pidx];
|
|
beta[bidx_ts] = std::max(beta[bidx_ts], T(NEGATIVE_CUTOFF_VAL));
|
|
}
|
|
|
|
return beta;
|
|
}
|
|
|
|
template <typename T>
|
|
void ctc_gradient_log(std::vector<int>& label,
|
|
const int input_length,
|
|
const int max_time_step,
|
|
const int batch_size,
|
|
const int class_sz,
|
|
const int pstr0,
|
|
const int pstr1,
|
|
const int pstr2,
|
|
const int gstr0,
|
|
const int gstr1,
|
|
const int gstr2,
|
|
const int batch_id,
|
|
std::vector<T>& probs_logits,
|
|
std::vector<T>& gradients_logits,
|
|
int blank_lb = 0)
|
|
{
|
|
int probs_stride[] = {pstr0, pstr1, pstr2};
|
|
int grads_stride[] = {gstr0, gstr1, gstr2};
|
|
int label_prime_len = 2 * label.size() + 1;
|
|
blank_lb = blank_lb < 0 ? 0 : (blank_lb >= class_sz ? class_sz - 1 : blank_lb);
|
|
std::vector<int> label_prime(label_prime_len, blank_lb);
|
|
for(int i = 0; i < label.size(); i++)
|
|
label_prime[2 * i + 1] = label[i];
|
|
|
|
std::vector<T> alpha_log = ctc_forward_log(probs_logits,
|
|
label,
|
|
input_length,
|
|
batch_size,
|
|
class_sz,
|
|
pstr0,
|
|
pstr1,
|
|
pstr2,
|
|
batch_id,
|
|
blank_lb);
|
|
std::vector<T> beta_log = ctc_backward_log(probs_logits,
|
|
label,
|
|
input_length,
|
|
batch_size,
|
|
class_sz,
|
|
pstr0,
|
|
pstr1,
|
|
pstr2,
|
|
batch_id,
|
|
blank_lb);
|
|
|
|
float prob_lx_log = logaddexp(alpha_log[alpha_log.size() - 1], alpha_log[alpha_log.size() - 2]);
|
|
|
|
for(int j = 0; j < input_length; j++)
|
|
for(int i = 0; i < class_sz; i++)
|
|
{
|
|
size_t pidx = j * probs_stride[0] + batch_id * probs_stride[1] + i;
|
|
size_t gidx = j * grads_stride[0] + batch_id * grads_stride[1] + i;
|
|
gradients_logits[gidx] = T(NEGATIVE_CUTOFF_VAL);
|
|
|
|
for(int k = 0; k < label_prime_len; k++)
|
|
if(label_prime[k] == i)
|
|
{
|
|
size_t kidx = j * label_prime_len + k;
|
|
gradients_logits[gidx] =
|
|
logaddexp(gradients_logits[gidx], alpha_log[kidx] + beta_log[kidx]);
|
|
}
|
|
gradients_logits[gidx] -= (2 * probs_logits[pidx]);
|
|
gradients_logits[gidx] -= prob_lx_log;
|
|
gradients_logits[gidx] = std::max(gradients_logits[gidx], T(NEGATIVE_CUTOFF_VAL));
|
|
gradients_logits[gidx] = -exp(gradients_logits[gidx]);
|
|
}
|
|
|
|
for(int j = input_length; j < max_time_step; j++)
|
|
for(int i = 0; i < class_sz; i++)
|
|
{
|
|
size_t gidx = j * grads_stride[0] + batch_id * grads_stride[1] + i;
|
|
gradients_logits[gidx] = 0;
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void ctc_softmaxlayer_gradient_log(std::vector<int>& label,
|
|
const int input_length,
|
|
const int batch_size,
|
|
const int class_sz,
|
|
const int pstr0,
|
|
const int pstr1,
|
|
const int pstr2,
|
|
const int gstr0,
|
|
const int gstr1,
|
|
const int gstr2,
|
|
const int batch_id,
|
|
std::vector<T>& probs_logits,
|
|
std::vector<T>& gradients_logits,
|
|
int blank_lb = 0)
|
|
{
|
|
int probs_stride[] = {pstr0, pstr1, pstr2};
|
|
int grads_stride[] = {gstr0, gstr1, gstr2};
|
|
int label_prime_len = 2 * label.size() + 1;
|
|
blank_lb = blank_lb < 0 ? 0 : (blank_lb >= class_sz ? class_sz - 1 : blank_lb);
|
|
std::vector<int> label_prime(label_prime_len, blank_lb);
|
|
for(int i = 0; i < label.size(); i++)
|
|
label_prime[2 * i + 1] = label[i];
|
|
|
|
std::vector<T> alpha_log = ctc_forward_log(probs_logits,
|
|
label,
|
|
input_length,
|
|
batch_size,
|
|
class_sz,
|
|
pstr0,
|
|
pstr1,
|
|
pstr2,
|
|
batch_id,
|
|
blank_lb);
|
|
std::vector<T> beta_log = ctc_backward_log(probs_logits,
|
|
label,
|
|
input_length,
|
|
batch_size,
|
|
class_sz,
|
|
pstr0,
|
|
pstr1,
|
|
pstr2,
|
|
batch_id,
|
|
blank_lb);
|
|
|
|
float prob_lx_log = logaddexp(alpha_log[alpha_log.size() - 1], alpha_log[alpha_log.size() - 2]);
|
|
|
|
for(int j = 0; j < input_length; j++)
|
|
for(int i = 0; i < class_sz; i++)
|
|
{
|
|
size_t pidx = j * probs_stride[0] + batch_id * probs_stride[1] + i;
|
|
size_t gidx = j * grads_stride[0] + batch_id * grads_stride[1] + i;
|
|
gradients_logits[gidx] = T(NEGATIVE_CUTOFF_VAL);
|
|
|
|
for(int k = 0; k < label_prime_len; k++)
|
|
if(label_prime[k] == i)
|
|
{
|
|
size_t kidx = j * label_prime_len + k;
|
|
gradients_logits[gidx] =
|
|
logaddexp(gradients_logits[gidx], alpha_log[kidx] + beta_log[kidx]);
|
|
}
|
|
gradients_logits[gidx] -= probs_logits[pidx];
|
|
gradients_logits[gidx] -= prob_lx_log;
|
|
gradients_logits[gidx] = std::max(gradients_logits[gidx], T(NEGATIVE_CUTOFF_VAL));
|
|
|
|
gradients_logits[gidx] = exp(probs_logits[pidx]) - exp(gradients_logits[gidx]);
|
|
}
|
|
}
|
|
|
|
template <typename Tgpu, typename Tref = Tgpu>
|
|
void RunCTCLossCPUVerify(const int num_class,
|
|
std::vector<size_t> probsSize,
|
|
std::vector<size_t> probsStride,
|
|
std::vector<size_t> gradientsSize,
|
|
std::vector<size_t> gradientsStride,
|
|
std::vector<Tgpu>& probs,
|
|
std::vector<int>& labels,
|
|
std::vector<int>& labelLengths,
|
|
std::vector<int>& inputLengths,
|
|
std::vector<Tref>& losses_host,
|
|
std::vector<Tref>& gradients_host,
|
|
std::vector<Tref>& workspace_host,
|
|
bool parallel,
|
|
const int blank_lb = 0,
|
|
bool is_softmax_applied = true,
|
|
const int verify_path = 1)
|
|
{
|
|
if(labelLengths.size() != inputLengths.size())
|
|
{
|
|
printf("Label batch size does not match input batch size\n");
|
|
return;
|
|
}
|
|
|
|
int class_sz = num_class + 1;
|
|
int max_time_step = *std::max_element(inputLengths.begin(), inputLengths.end());
|
|
int batch_size = inputLengths.size();
|
|
|
|
if(probs.size() != (max_time_step * batch_size * class_sz))
|
|
{
|
|
printf("Wrong probability tensor size\n");
|
|
return;
|
|
}
|
|
if(probs.size() != gradients_host.size())
|
|
{
|
|
printf("Wrong gradient tensor size\n");
|
|
return;
|
|
}
|
|
if(probsSize[0] != max_time_step || probsSize[1] != batch_size || probsSize[2] != class_sz ||
|
|
gradientsSize[0] != max_time_step || gradientsSize[1] != batch_size ||
|
|
gradientsSize[2] != class_sz)
|
|
{
|
|
printf("Wrong tensor size\n");
|
|
return;
|
|
}
|
|
|
|
std::vector<Tref> beta_loss(batch_size, 0);
|
|
if(verify_path == 1)
|
|
{
|
|
std::vector<int> probsDesc = {max_time_step,
|
|
batch_size,
|
|
class_sz,
|
|
int(probsStride[0]),
|
|
int(probsStride[1]),
|
|
int(probsStride[2])};
|
|
std::vector<int> gradientsDesc = {max_time_step,
|
|
batch_size,
|
|
class_sz,
|
|
int(gradientsStride[0]),
|
|
int(gradientsStride[1]),
|
|
int(gradientsStride[2])};
|
|
|
|
RunCTCLossGPUEmulator(probsDesc,
|
|
probs,
|
|
labels.data(),
|
|
labelLengths.data(),
|
|
inputLengths.data(),
|
|
losses_host,
|
|
gradientsDesc,
|
|
gradients_host,
|
|
workspace_host,
|
|
beta_loss,
|
|
parallel,
|
|
blank_lb,
|
|
is_softmax_applied);
|
|
}
|
|
else
|
|
{
|
|
std::vector<Tref> probs_logits(probs.size(), Tref(NEGATIVE_CUTOFF_VAL));
|
|
std::vector<Tref> gradients_logits(probs.size(), Tref(NEGATIVE_CUTOFF_VAL));
|
|
std::vector<Tref> softmaxlayer_gradients_logit(probs.size(), 0);
|
|
|
|
for(int j = 0; j < max_time_step * batch_size; j++)
|
|
subvec_logsoftmax(probs, probs_logits, j * class_sz, j * class_sz, class_sz);
|
|
|
|
auto probs_logits_use = is_softmax_applied ? probs_logits : probs;
|
|
|
|
std::vector<int> label_offsets(batch_size, 0);
|
|
for(int j = 1; j < batch_size; j++)
|
|
label_offsets[j] = label_offsets[j - 1] + labelLengths[j - 1];
|
|
|
|
auto per_batch_work = [&](int j) {
|
|
auto lab_begin = labels.begin() + label_offsets[j];
|
|
std::vector<int> indiv_lab(lab_begin, lab_begin + labelLengths[j]);
|
|
|
|
std::vector<Tref> alpha_log = ctc_forward_log(probs_logits_use,
|
|
indiv_lab,
|
|
inputLengths[j],
|
|
batch_size,
|
|
class_sz,
|
|
int(probsStride[0]),
|
|
int(probsStride[1]),
|
|
int(probsStride[2]),
|
|
j,
|
|
blank_lb);
|
|
|
|
float losses_log =
|
|
-logaddexp(alpha_log[alpha_log.size() - 1], alpha_log[alpha_log.size() - 2]);
|
|
losses_host[j] = losses_log;
|
|
|
|
std::vector<Tref> beta_log = ctc_backward_log(probs_logits_use,
|
|
indiv_lab,
|
|
inputLengths[j],
|
|
batch_size,
|
|
class_sz,
|
|
int(probsStride[0]),
|
|
int(probsStride[1]),
|
|
int(probsStride[2]),
|
|
j,
|
|
blank_lb);
|
|
|
|
beta_loss[j] = logaddexp(beta_log[0], beta_log[1]);
|
|
|
|
if(is_softmax_applied)
|
|
ctc_softmaxlayer_gradient_log(indiv_lab,
|
|
inputLengths[j],
|
|
batch_size,
|
|
class_sz,
|
|
int(probsStride[0]),
|
|
int(probsStride[1]),
|
|
int(probsStride[2]),
|
|
int(gradientsStride[0]),
|
|
int(gradientsStride[1]),
|
|
int(gradientsStride[2]),
|
|
j,
|
|
probs_logits_use,
|
|
softmaxlayer_gradients_logit,
|
|
blank_lb);
|
|
else
|
|
ctc_gradient_log(indiv_lab,
|
|
inputLengths[j],
|
|
max_time_step,
|
|
batch_size,
|
|
class_sz,
|
|
int(probsStride[0]),
|
|
int(probsStride[1]),
|
|
int(probsStride[2]),
|
|
int(gradientsStride[0]),
|
|
int(gradientsStride[1]),
|
|
int(gradientsStride[2]),
|
|
j,
|
|
probs_logits_use,
|
|
gradients_logits,
|
|
blank_lb);
|
|
};
|
|
|
|
if(parallel)
|
|
{
|
|
miopen::par_for(batch_size, per_batch_work);
|
|
}
|
|
else
|
|
{
|
|
for(int j = 0; j < batch_size; j++)
|
|
{
|
|
per_batch_work(j);
|
|
}
|
|
}
|
|
|
|
gradients_host = is_softmax_applied ? softmaxlayer_gradients_logit : gradients_logits;
|
|
|
|
(void)probsStride;
|
|
(void)gradientsStride;
|
|
(void)workspace_host;
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void GetCTCLossWorkspaceSizeCPU(std::vector<size_t> probsDesc,
|
|
std::vector<size_t> gradientsDesc,
|
|
const int* labels,
|
|
const int* labelLengths,
|
|
const int* inputLengths,
|
|
miopenCTCLossAlgo_t ctc_algo,
|
|
size_t* workSpaceSizeCPU)
|
|
{
|
|
(void)ctc_algo;
|
|
|
|
if(probsDesc[0] != gradientsDesc[0] || probsDesc[1] != gradientsDesc[1] ||
|
|
probsDesc[2] != gradientsDesc[2])
|
|
{
|
|
printf("Label batch size does not match input batch size\n");
|
|
return;
|
|
}
|
|
|
|
int class_sz = probsDesc[2];
|
|
int batch_size = probsDesc[1];
|
|
int max_time_step = probsDesc[0];
|
|
int max_label_len = 0;
|
|
int total_label_len = 0;
|
|
std::vector<int> repeat(batch_size, 0);
|
|
std::vector<int> labels_offset(batch_size, 0);
|
|
size_t wksp_sz_lb = 0;
|
|
size_t wksp_sz_dat = 0;
|
|
|
|
for(int i = 0; i < batch_size; i++)
|
|
{
|
|
if(inputLengths[i] > max_time_step)
|
|
{
|
|
printf("Wrong input time step\n");
|
|
return;
|
|
}
|
|
max_label_len = std::max(max_label_len, labelLengths[i]);
|
|
total_label_len += labelLengths[i];
|
|
labels_offset[i] = i == 0 ? 0 : (labels_offset[i - 1] + labelLengths[i - 1]);
|
|
|
|
for(int j = 0; j < labelLengths[i]; j++)
|
|
{
|
|
if(labels[labels_offset[i] + j] >= class_sz)
|
|
{
|
|
printf("Wrong label id at batch\n");
|
|
return;
|
|
}
|
|
if(j > 0)
|
|
if(labels[labels_offset[i] + j] == labels[labels_offset[i] + j - 1])
|
|
repeat[i]++;
|
|
}
|
|
|
|
if(labelLengths[i] + repeat[i] > inputLengths[i])
|
|
{
|
|
printf("Error: label length exceeds input time step\n");
|
|
return;
|
|
}
|
|
}
|
|
|
|
// input length
|
|
wksp_sz_lb += batch_size;
|
|
|
|
// label length
|
|
wksp_sz_lb += batch_size;
|
|
|
|
// label offset
|
|
wksp_sz_lb += batch_size;
|
|
|
|
// label repeat
|
|
wksp_sz_lb += batch_size;
|
|
|
|
// labels
|
|
wksp_sz_lb += total_label_len;
|
|
|
|
// labels with blanks
|
|
wksp_sz_lb += batch_size * (2 * max_label_len + 1);
|
|
|
|
// logsoftmax of probs
|
|
wksp_sz_dat += max_time_step * batch_size * class_sz;
|
|
|
|
// alphas
|
|
wksp_sz_dat += max_time_step * batch_size * (2 * max_label_len + 1);
|
|
|
|
*workSpaceSizeCPU = (wksp_sz_dat + wksp_sz_lb) * sizeof(T);
|
|
}
|