|
| 1 | +// Copyright (C) 2024 Intel Corporation |
| 2 | +// SPDX-License-Identifier: Apache-2.0 |
| 3 | +// |
| 4 | + |
| 5 | +#include "include/batch_headers/common.cl" |
| 6 | +#include "include/batch_headers/sub_group_block_read.cl" |
| 7 | +#include "include/batch_headers/sub_group_block_write.cl" |
| 8 | +#include "include/batch_headers/sub_group_shuffle.cl" |
| 9 | + |
| 10 | + |
| 11 | + |
| 12 | + |
| 13 | + |
| 14 | +// constexpr size_t HEAD_SIZE = 64; |
| 15 | +// constexpr size_t HEADS_NUM = 32; |
| 16 | +// constexpr size_t KV_HEADS_NUM = 4; |
| 17 | +// constexpr size_t BLOCK_SIZE = 16; |
| 18 | +// constexpr size_t X_SIZE = 4; |
| 19 | + |
| 20 | +// constexpr size_t MAX_SEQUENCE_LENGTH = 1024; |
| 21 | + |
| 22 | + |
| 23 | + |
| 24 | +#define SUB_GROUP_SIZE 16 |
| 25 | + |
| 26 | +// The size of portion of HEAD_SIZE each WI process |
| 27 | +#define HEAD_ITEMS_PER_WI (HEAD_SIZE / SUB_GROUP_SIZE) |
| 28 | + |
| 29 | +// How much QK outputs each subgroup calculates per cycle |
| 30 | +#define QK_PER_SG 4 |
| 31 | + |
| 32 | +#define KV_CACHE_BLOCK_STRIDE (HEAD_SIZE * HEADS_NUM * BLOCK_SIZE) |
| 33 | + |
| 34 | +#define QUERY_BLOCK_READ(ptr, offset) BLOCK_READN(INPUT0_TYPE, 1, ptr, offset) |
| 35 | + |
| 36 | +#define SUBGROUPS_PER_WG HEAD_SIZE / SUB_GROUP_SIZE |
| 37 | + |
| 38 | +REQD_SUB_GROUP_SIZE(SUB_GROUP_SIZE) |
| 39 | +__attribute__((reqd_work_group_size(1, 1, SUB_GROUP_SIZE))) |
| 40 | +KERNEL(pa_sdpa_ref)( |
| 41 | + OPTIONAL_SHAPE_INFO_ARG |
| 42 | + __global const INPUT0_TYPE* query, |
| 43 | + __global const INPUT1_TYPE* key_cache, |
| 44 | + __global const INPUT2_TYPE* value_cache, |
| 45 | + __global const INPUT3_TYPE* max_context_len, |
| 46 | + __global const INPUT4_TYPE* context_lens, |
| 47 | + __global const INPUT5_TYPE* block_tables, |
| 48 | + __global OUTPUT_TYPE* output) |
| 49 | +{ |
| 50 | + const uint seq_idx = get_global_id(0); |
| 51 | + const uint head_num_idx = get_global_id(1); |
| 52 | + const uint head_idx = get_global_id(2); |
| 53 | + const uint sglid = get_sub_group_local_id(); |
| 54 | + const uint sgid = get_sub_group_id(); |
| 55 | + |
| 56 | + const uint batch_idx = seq_idx / INPUT0_FEATURE_NUM; |
| 57 | + const uint token_idx = seq_idx % INPUT0_FEATURE_NUM; |
| 58 | + |
| 59 | + const uint context_len = context_lens[batch_idx]; |
| 60 | + |
| 61 | + const uint blocks_num = INPUT5_FEATURE_NUM; |
| 62 | + |
| 63 | + // sgid0: 0..3 |
| 64 | + // sgid1: 4..7 |
| 65 | + // sgid2: 8..11 |
| 66 | + // sgid3: 12..15 |
| 67 | + |
| 68 | + // sgid0: 16..19 |
| 69 | + // sgid1: 20..23 |
| 70 | + // sgid2: 24..27 |
| 71 | + // sgid3: 28..31 |
| 72 | + |
| 73 | + // TODO: Need to make blocks division more flexible. Current approach suggests |
| 74 | + // to have 4 SG per WG, where each SG process 4 QK outputs, so 16 in total per WG |
| 75 | + |
| 76 | + __local OUTPUT_TYPE qk_vals[SHARED_MEM_SIZE]; |
| 77 | + |
| 78 | + OUTPUT_TYPE qk_max = OUTPUT_VAL_MIN; |
| 79 | + |
| 80 | + for (uint block = 0; block < blocks_num; block++) { |
| 81 | + const uint block_idx = batch_idx * blocks_num + block; |
| 82 | + const uint block_offset = block_tables[block_idx] * KV_CACHE_BLOCK_STRIDE; |
| 83 | + |
| 84 | + OUTPUT_TYPE qk[QK_PER_SG] = {0}; |
| 85 | + |
| 86 | + for (uint hs = 0; hs < HEAD_ITEMS_PER_WI; hs++) { |
| 87 | + const uint query_idx = seq_idx * HEAD_SIZE * HEADS_NUM + hs * SUB_GROUP_SIZE; |
| 88 | + |
| 89 | + // TODO: can be preloaded outside HEAD_ITEMS_PER_WI loop - need to check perf |
| 90 | + INPUT0_TYPE q = QUERY_BLOCK_READ(query, query_idx); |
| 91 | + for (uint qk_idx = 0; qk_idx < QK_PER_SG; qk_idx++) { |
| 92 | + uint current_token = block * BLOCK_SIZE + sgid * QK_PER_SG + qk_idx; |
| 93 | + if (current_token >= context_len) |
| 94 | + continue; |
| 95 | + |
| 96 | + const uint key_idx = block_offset + |
| 97 | + (X_SIZE * QK_PER_SG) * sgid + |
| 98 | + (HEAD_ITEMS_PER_WI * BLOCK_SIZE * X_SIZE) * hs + |
| 99 | + (sglid / X_SIZE) * X_SIZE * BLOCK_SIZE + |
| 100 | + (sglid % X_SIZE) + qk_idx * X_SIZE; |
| 101 | + // TODO1: try block loading and shuffling |
| 102 | + // TODO2: try to load k*4 times and then calculate |
| 103 | + // TODO3: try bigger X block |
| 104 | + INPUT1_TYPE k = key_cache[key_idx]; |
| 105 | + |
| 106 | + qk[qk_idx] = mad(q, k, qk[qk_idx]); |
| 107 | + } |
| 108 | + } |
| 109 | + |
| 110 | + // Summurize qk calculation across all WIs |
| 111 | + for (uint qk_idx = 0; qk_idx < QK_PER_SG; qk_idx++) { |
| 112 | + qk[QK_PER_SG] = sub_group_reduce_add(qk[QK_PER_SG]); |
| 113 | + qk_max = OUTPUT_MAX_FUNC(qk_max, qk[QK_PER_SG]); |
| 114 | + } |
| 115 | + |
| 116 | + // Save QK results to local memory |
| 117 | + if (sglid < QK_PER_SG) { |
| 118 | + const uint qk_local_idx = block * BLOCK_SIZE * sgid * QK_PER_SG + sglid; |
| 119 | + qk_vals[qk_local_idx] = qk[sglid]; |
| 120 | + } |
| 121 | + } |
| 122 | + |
| 123 | + /* WARNING NEED TO ADD BIAS BEFORE SOFTMAX */ |
| 124 | + |
| 125 | + // Apply SoftMax operation |
| 126 | + __local OUTPUT_TYPE qk_max_vals[SUBGROUPS_PER_WG]; |
| 127 | + __local OUTPUT_TYPE qk_sum_vals[SUBGROUPS_PER_WG]; |
| 128 | + { |
| 129 | + if (sglid == 0) |
| 130 | + qk_max_vals[sgid] = qk_max; |
| 131 | + |
| 132 | + barrier(CLK_LOCAL_MEM_FENCE); |
| 133 | + |
| 134 | + qk_max = OUTPUT_VAL_MIN; |
| 135 | + if (sglid < SUBGROUPS_PER_WG) |
| 136 | + qk_max = qk_max_vals[sglid]; |
| 137 | + |
| 138 | + // Final max value after reduction across of all SG and WI |
| 139 | + qk_max = sub_group_reduce_max(qk_max); |
| 140 | + |
| 141 | + OUTPUT_TYPE exp_sum = OUTPUT_VAL_ZERO; |
| 142 | + for (uint qk_idx = 0; qk_idx < CEIL_DIV(context_len, SUBGROUPS_PER_WG * SUB_GROUP_SIZE); qk_idx++) { |
| 143 | + const uint data_idx = qk_idx * (SUBGROUPS_PER_WG * SUB_GROUP_SIZE) + sgid * SUB_GROUP_SIZE + sglid; |
| 144 | + if (data_idx < context_len) { |
| 145 | + OUTPUT_TYPE val = native_exp(qk_vals[data_idx] - qk_max); |
| 146 | + exp_sum += val; |
| 147 | + qk_vals[data_idx] = val; |
| 148 | + } |
| 149 | + } |
| 150 | + |
| 151 | + exp_sum = sub_group_reduce_add(exp_sum); |
| 152 | + |
| 153 | + if (sglid == 0) |
| 154 | + qk_sum_vals[sgid] = exp_sum; |
| 155 | + |
| 156 | + barrier(CLK_LOCAL_MEM_FENCE); |
| 157 | + |
| 158 | + if (sglid < SUBGROUPS_PER_WG) |
| 159 | + exp_sum = qk_sum_vals[sglid]; |
| 160 | + |
| 161 | + // Final sum of all values |
| 162 | + exp_sum = sub_group_reduce_add(exp_sum); |
| 163 | + |
| 164 | + const OUTPUT_TYPE inv_sum = OUTPUT_VAL_ONE / exp_sum; |
| 165 | + |
| 166 | + for (uint qk_idx = 0; qk_idx < CEIL_DIV(context_len, SUBGROUPS_PER_WG * SUB_GROUP_SIZE); qk_idx++) { |
| 167 | + const uint data_idx = qk_idx * (SUBGROUPS_PER_WG * SUB_GROUP_SIZE) + sgid * SUB_GROUP_SIZE + sglid; |
| 168 | + if (data_idx < context_len) { |
| 169 | + OUTPUT_TYPE val = qk_vals[data_idx] * inv_sum; |
| 170 | + qk_vals[data_idx] = val; |
| 171 | + } |
| 172 | + } |
| 173 | + |
| 174 | + barrier(CLK_LOCAL_MEM_FENCE); |
| 175 | + } |
| 176 | + |
| 177 | + output[seq_idx + sglid] = qk_vals[sglid % context_len]; |
| 178 | +} |
0 commit comments