ctc_include/detail/gpu_ctc_kernels.h

/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
 *
 *   http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
 * KIND, either express or implied.  See the License for the
 * specific language governing permissions and limitations
 * under the License.
 */

#pragma once

#include "../contrib/moderngpu/include/device/ctascan.cuh"
#include "../contrib/moderngpu/include/device/ctamerge.cuh"

#include "ctc_helper.h"

using namespace mgpu;

template<int NT, int VT, typename T, typename KeyT, typename Op>
struct CTASegReduce {

    enum {NV = NT * VT};

    union Storage {
        typename CTAScan<NT>::Storage scanStorage;
        int indices[NV];
    };

    //adapted from global kernel KernelReduceByKeyPreprocess
    __device__ static void preprocessKeys(KeyT *keys, int count,
                                          int *numUniqueLabels, int seg_start[VT],
                                          int seg_end[VT], int *scanout) {
        __shared__ Storage shared;

        const int tid = threadIdx.x;
        // Compare adjacent keys within each thread and mark discontinuities
        int endFlags = 0;
        T key = keys[VT * tid];
        #pragma unroll
        for (int i = 0; i < VT; ++i) {
            int index = VT * tid + 1 + i;
            T next = keys[index];
            if(index == count || (index < count && key != next)) {
                endFlags |= 1 << i;
            }
            key = next;
        }

        __syncthreads();

        //Count the number of encountered end flags
        int scan = CTAScan<NT>::Scan(tid, popc(endFlags), shared.scanStorage, numUniqueLabels);

        __syncthreads();

        //output the unique keys
        //use indices as scratch space
        int outputPos = scan;
        #pragma unroll
        for (int i = 0; i < VT; ++i) {

            if ( (endFlags >> i) & 1) {
                shared.indices[outputPos] = keys[VT * tid + i];
                scanout[outputPos] = VT * tid + i;
                outputPos++;
            }
        }

        __syncthreads();

        // Create start and end
        for (int idx = tid, j = 0; idx < (*numUniqueLabels); idx += blockDim.x, ++j) {
            seg_start[j] = (idx == 0) ? 0 : (scanout[idx-1] + 1);
            seg_end[j] = scanout[idx];
        }

        __syncthreads();

        //copy from the scratch space back into the keys
        #pragma unroll
        for (int i = 0; i < VT; ++i) {
            keys[i * NT + tid] = shared.indices[i * NT + tid];
        }

        __syncthreads();
    }
};

// Computes forward probabilities. This fills in a T * S matrix.
// The computation starts at t=1 (2nd row) and ends at t=T-1 (last row). Each row has
// S elements where S = 2L + 1.
//
// We only need to read in probabilities corresponding to the labels, thus a sparse
// set of values are read from the log probs matrix since the character set is much smaller
// than the labels. This is much more true for Mandarin than English.
template<typename ProbT, int NT, int VT>
__global__
void compute_alpha_kernel (const ProbT* log_probs, const int *label_sizes,
                           const int *utt_length, const int *repeats_in_labels,
                           const int *labels_without_blanks, const int *label_offsets,
                           int *labels_with_blanks, ProbT *alphas,
                           ProbT* nll_forward, int stride, int out_dim,
                           int S_memoffset, int T_memoffset, int blank_label) {

    ctc_helper::log_plus<ProbT> log_plus_f;

    const int tid = threadIdx.x;
    const int L = label_sizes[blockIdx.x];
    const int T = utt_length[blockIdx.x];
    const int S = 2*L + 1;
    const int prob_offset = out_dim * blockIdx.x;
    const int repeats = repeats_in_labels[blockIdx.x];

    const int NV = NT * VT;
    __shared__ int label[NV];

    if ((L + repeats) > T)
        return;

    // Generate labels with blanks from labels without blanks
    {
        const int label_start_offset = label_offsets[blockIdx.x];
        for (int idx = tid; idx < L; idx += blockDim.x) {
            const int offset = (blockIdx.x * S_memoffset) + 2 * idx;
            labels_with_blanks[offset] = blank_label;
            labels_with_blanks[offset+1] = labels_without_blanks[label_start_offset + idx];
        }
        if (tid == 0) {
            labels_with_blanks[(blockIdx.x * S_memoffset) + 2 * L] = blank_label;
        }
    }
    __syncthreads();

    const int *labels = labels_with_blanks;
    const int* label_global = &labels[blockIdx.x * S_memoffset];
    ProbT* alpha = &alphas[blockIdx.x * (S_memoffset * T_memoffset)];

    // Set the first row of alpha neg_inf - it is much more efficient to do it
    // here than outside
    #pragma unroll
    for (int idx = tid; idx < min(S, NV); idx += blockDim.x) {
        alpha[idx] = ctc_helper::neg_inf<ProbT>();
    }

    // Load labels into shared memory
    #pragma unroll
    for (int i = tid; i < S; i += NT) {
        label[i] = label_global[i];
    }

    __syncthreads();

    int start =  (L + repeats < T) ? 0 : 1;
    int end = S > 1 ? 2 : 1;

    // Initialize the first row corresponding to t=0;
    for(int i = tid; i < (end-start); i += blockDim.x)
        alpha[i + start] = log_probs[prob_offset + label[i + start]];

    __syncthreads();

    // Fill in the rest of matrix, one row at a time (outer loop).
    for(int t = 1; t < T; ++t) {

        // Start offsets into the current and previous row
        const int start_cur_row = t * S;
        const int start_prev_row = (t - 1) * S;

        // The prob is a 2D column major array, with probabilites for each t strided
        // by (out_dim * stride), where stride is the minibatch size
        const int start_prob_col = t * (out_dim * stride);

        // This is the first column and in this case there is nothing left of it
        if (tid == 0) {
            if (start == 0) {
                alpha[start_cur_row] = alpha[start_prev_row] +
                                       log_probs[prob_offset + start_prob_col + blank_label];
            }
            else if (start == 1) {
                alpha[start_cur_row] = alpha[start_prev_row];
            }
        }

        __syncthreads();

        // Fill in the elements in each row. There is no loop dependence here since our
        // input is the row above. We sum either two or three adjacent values from the
        // row above depending on whether we have a blank or repeated characters. Finally
        // we add the probability corresponding to this label at time t
        #pragma unroll
        for (int idx = (tid+1); idx < S; idx += blockDim.x) {

            ProbT prev_sum = log_plus_f(alpha[idx + start_prev_row], alpha[(idx-1) + start_prev_row]);

            // Skip two if not on blank and not on repeat.
            if ((label[idx] != blank_label) &&
                (idx != 1) && (label[idx] != label[idx-2]))
                prev_sum = log_plus_f(prev_sum, alpha[(idx-2) + start_prev_row]);

            alpha[idx + start_cur_row] =
                prev_sum + log_probs[prob_offset + start_prob_col + label[idx]];
        }

        __syncthreads();
    }

    if (tid == 0) {
        // Add and return the rightmost two/one element(s) in the last row.
        ProbT loglike = ctc_helper::neg_inf<ProbT>();

        // This is the total increment for s_inc and e_inc through the loop
        const int val = 2 * (L-1) + 1 - (((L + repeats) == T) ? 1 : 0);

        start = (val * (L!=0) + start);
        end = (val * (L!=0) + end);

        for(int i = start; i < end; ++i)
            loglike = log_plus_f(loglike, alpha[i + (T - 1) * S]);

        nll_forward[blockIdx.x] = -loglike;
    }
}

// Computes backward probabilities. This also fills in a T * S matrix
//
// See comments above compute_alphas for more context.
template<typename ProbT, int NT, int VT>
__global__
void compute_betas_and_grad_kernel (const ProbT* log_probs, const int *label_sizes,
                                    const int *utt_length, const int *repeats_in_labels,
                                    const int *labels_with_blanks, ProbT *alphas,
                                    const ProbT* nll_forward, ProbT *nll_backward,
                                    ProbT *grads, int stride, int out_dim,
                                    int S_memoffset, int T_memoffset, int blank_label) {

    ctc_helper::log_plus<ProbT> log_plus_f;
    typedef CTASegReduce<NT, VT, ProbT, int, ctc_helper::log_plus<ProbT>> SegReduce;

    const int tid = threadIdx.x;
    const int L = label_sizes[blockIdx.x];
    const int T = utt_length[blockIdx.x];
    const int S = 2*L + 1;
    const int prob_offset = out_dim * blockIdx.x;
    const int repeats = repeats_in_labels[blockIdx.x];
    const ProbT log_partition = -nll_forward[blockIdx.x];

    const int* labels = labels_with_blanks;
    const int* label_global = &labels[blockIdx.x * S_memoffset];
    ProbT* alpha = &alphas[blockIdx.x * (S_memoffset * T_memoffset)];

    const int NV = NT * VT;

    union TempStorage {
        ProbT beta[NV];
        int result[NV];
    };

    __shared__ TempStorage temp_buffer;

    __shared__ int label[NV];

    // Temporaries needed for segmented reduce
    // TODO: see if we can combine the shared memory requirements
    __shared__ int keys_shared[NV];
    __shared__ int gather_indices[NV];
    __shared__ ProbT output[NV];

    ProbT beta_val[VT];

    if ((L + repeats) > T)
        return;

    int start = S > 1 ? (S - 2) : 0;
    int end = (L + repeats < T) ? S : S-1;

    // Setup shared memory buffers
    #pragma unroll
    for (int idx = tid; idx < NV; idx += NT) {
        label[idx] = (idx < S) ? label_global[idx] : INT_MAX;
    }

    __syncthreads();

    // int flags;
    int uniquelabels;
    int seg_start[VT];
    int seg_end[VT];

    // Sort labels and record indices from which to gather from
    {
        int key[VT];
        int gather_val[VT];

        #pragma unroll
        for (int i = 0; i < VT; ++i) {
            const int idx = tid * VT + i;
            gather_val[i] = idx;
            key[i] = label[idx];
        }

        __syncthreads();

        CTAMergesort<NT, VT, true, true, int, int, mgpu::less<int>>
            (key, gather_val, keys_shared, gather_indices, S, tid, mgpu::less<int>());

        __syncthreads();

        for (int i = 0; i < VT; ++i) {
            const int idx = tid * VT + i;
            gather_indices[idx] = gather_val[i];
        }

        __syncthreads();

        SegReduce::preprocessKeys(keys_shared, S, &uniquelabels, seg_start, seg_end,
                                  temp_buffer.result);
        __syncthreads();
    }

    // TODO: probably not necessary
    __syncthreads();

    // Load labels back
    #pragma unroll
    for (int idx = tid; idx < NV; idx += NT) {
        temp_buffer.beta[idx] = ctc_helper::neg_inf<ProbT>();
    }
    __syncthreads();

    // Initialize the two rightmost values in the last row (assuming L non-zero)
    for(int i = tid; i < (end-start); i += blockDim.x)
        temp_buffer.beta[i + start] =
            log_probs[prob_offset + (T - 1) * (out_dim * stride) + label[i + start]];

    __syncthreads();

    // Load output data in registers through the transpose trick - should really be a function
    #pragma unroll
    for (int idx = tid; idx < S; idx += NT) {
        output[idx] = alpha[idx + (T - 1) * S] + temp_buffer.beta[idx];
    }

    __syncthreads();

    // Start at the second to last row and backward in time
    for(int t = T - 1; t >= 0; --t) {

        // Start offsets into the current and next row
        const int start_cur_row = t * S;

        // Starting offset of column that we read from the log probs array
        const int start_prob_col = t * (out_dim * stride);

        if (t < T-1) {

            // Filling up one row at at time but going back in time from the last row
            // to the first. As in the forward pass, there is no loop dependence and we
            // do a variable length filter of maximum filter size of 3
            #pragma unroll
            for(int idx = tid, i = 0; idx < (S-1); idx += NT, i++) {
                ProbT next_sum = log_plus_f(temp_buffer.beta[idx], temp_buffer.beta[idx+1]);

                    // Skip two if not on blank and not on repeat.
                if ((label[idx] != blank_label) &&
                    (idx != (S-2)) && (label[idx] != label[idx+2]))
                    next_sum = log_plus_f(next_sum, temp_buffer.beta[idx+2]);

                beta_val[i] = next_sum + log_probs[prob_offset + start_prob_col + label[idx]];
            }

            __syncthreads();

            // Initialize values for the rightmost column since there is nothing to the right
            // Update input buffer for next iteration
            if ((tid == 0) && (end == S))
                temp_buffer.beta[(S-1)] = temp_buffer.beta[(S-1)] +
                                          log_probs[prob_offset + start_prob_col + blank_label];

            #pragma unroll
            for(int idx = tid, i = 0; idx < (S-1); idx += NT, i++) {
               temp_buffer.beta[idx] = beta_val[i];
            }

            __syncthreads();

            // Beta Computation done - add to alpha and update the gradient. Reload
            // the gradient back for segmented reduce later on
            #pragma unroll
            for(int idx = tid; idx < S; idx += NT) {
               output[idx] = alpha[idx + start_cur_row] + temp_buffer.beta[idx];
            }

            __syncthreads();

        }

        __syncthreads();

        // Compute segmented reduction of output by using label as key
        {
            // Somewhat faster key value reduce
            ProbT accum[VT];

            for (int idx = tid, j = 0; idx < uniquelabels; idx += blockDim.x, ++j) {

                accum[j] = ctc_helper::neg_inf<ProbT>();
                for (int i = seg_start[j]; i <= seg_end[j]; ++i) {
                    accum[j] = log_plus_f(accum[j], output[gather_indices[i]]);
                }
            }
            __syncthreads();

            // Write accumulated value into output since that is not used
            for (int idx = tid, j = 0; idx < uniquelabels; idx += blockDim.x, ++j) {
                output[idx] = accum[j];
            }
            __syncthreads();

            for (int idx = tid; idx < out_dim; idx += blockDim.x) {
                const int grads_offset = prob_offset + start_prob_col + idx;
                grads[grads_offset] = exp(log_probs[grads_offset]);
            }

            __syncthreads();

            for (int idx = tid; idx < uniquelabels; idx += blockDim.x) {
                const int grads_offset = prob_offset + start_prob_col + keys_shared[idx];

                ProbT grad = output[idx];

                if ((grad == 0.0) || (log_probs[grads_offset] == ctc_helper::neg_inf<ProbT>()) ||
                    (grad == ctc_helper::neg_inf<ProbT>())) {
                } else {
                    grads[grads_offset] =
                        exp(log_probs[grads_offset]) - exp(grad - log_probs[grads_offset] - log_partition);
                }
            }

            __syncthreads();
        }

        // Output backward log likelihood
        if ((t == 0) && (tid == 0)) {
            ProbT loglike = ctc_helper::neg_inf<ProbT>();

            const int val = 2 * (L-1) + 1 - (((L + repeats) == T) ? 1 : 0);

            start = (-val * (L != 0) + start);
            end = (-val * (L != 0) + end);

            // Sum and return the leftmost one/two value(s) in first row
            for(int i = start; i < end; ++i)
                loglike = log_plus_f(loglike, temp_buffer.beta[i]);

            nll_backward[blockIdx.x] = -loglike;
        }

        // For some reason this is important
        __syncthreads();
    }
}

template <typename ProbT, int VT = 1, typename Op>
__global__ void compute_log_probs_kernel(Op f, ProbT* log_probs,
                                     const ProbT* const denom,
                                     int alphabet_size,
                                     int count) {

    int idx = blockDim.x * blockIdx.x + threadIdx.x;
    int stride = blockDim.x * gridDim.x;
#pragma unroll
    for(int i = 0; i < VT; i++) {
        if (idx < count) {
            const int column_idx = idx / alphabet_size;
            log_probs[idx] = log_probs[idx] - log(denom[column_idx]);
        }
        idx += stride;
    }
}

template <typename ProbT, int VT = 1, typename Op>
__global__ void prepare_stable_LSM_kernel(Op f, ProbT* log_probs,
                                         const ProbT* const col_max,
                                         int alphabet_size,
                                         int count) {

    int idx = blockDim.x * blockIdx.x + threadIdx.x;
    int stride = blockDim.x * gridDim.x;
#pragma unroll
    for(int i = 0; i < VT; i++) {
        if (idx < count) {
            const int column_idx = idx / alphabet_size;
            log_probs[idx] = f(log_probs[idx] - col_max[column_idx]);
        }
        idx += stride;
    }
}