1#pragma OPENCL EXTENSION cl_khr_fp16 : enable
  2
  3#define ACC_TYPE float
  4#define ACC_TYPE4 float4
  5#define DATA_TYPE half
  6#define DATA_TYPE4 half4
  7#define CONVERT_ACC4(x) convert_float4(x)
  8#define CONVERT_DATA4(x) convert_half4(x)
  9
 10#define DK_VEC (DK/4)
 11#define DV_VEC (DV/4)
 12#define WG_SIZE (BLOCK_M)
 13#define Q1_WG_SIZE 64
 14
 15inline float get_alibi_slope(
 16    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
 17) {
 18    if (max_bias <= 0.0f) {
 19        return 1.0f;
 20    }
 21    const float base = h < n_head_log2 ? m0 : m1;
 22    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
 23
 24    return pow(base, exph);
 25}
 26__kernel void flash_attn_f16(
 27    const global void * q_void, ulong q_offset,
 28    const global void * k_void, ulong k_offset,
 29    const global void * v_void, ulong v_offset,
 30    global void * o_void, ulong o_offset,
 31    const float scale,
 32    const int n_q,
 33    const int n_kv,
 34    const int is_causal,
 35    const int n_head,
 36    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
 37    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
 38    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
 39    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
 40    const float max_bias,
 41    const float m0,
 42    const float m1,
 43    const int n_head_log2,
 44    const float logit_softcap,
 45    const int n_head_kv,
 46    const global void* mask_void,
 47    const ulong mask_offset,
 48    const ulong mask_nb1,
 49    const ulong mask_nb2,
 50    const ulong mask_nb3,
 51    const int mask_ne2,
 52    const int mask_ne3,
 53    const global void* sinks_void,
 54    const ulong sinks_offset
 55) {
 56    const int tid = get_local_id(0);
 57    const int block_q_idx = get_group_id(0);
 58    const int head_batch_idx = get_global_id(1);
 59
 60    const int my_query_row = block_q_idx * BLOCK_M + tid;
 61
 62    const int batch_idx = head_batch_idx / n_head;
 63    const int head_idx = head_batch_idx % n_head;
 64
 65    const int gqa_ratio = n_head / n_head_kv;
 66    const int head_kv_idx = head_idx / gqa_ratio;
 67
 68    const global char* q_base = (const global char*)q_void + q_offset;
 69    const global char* k_base = (const global char*)k_void + k_offset;
 70    const global char* v_base = (const global char*)v_void + v_offset;
 71    global char* o_base = (global char*)o_void + o_offset;
 72
 73    const global char* mask_base = NULL;
 74    if (mask_void != NULL) {
 75        const int mask_head_idx = head_idx % mask_ne2;
 76        const int mask_batch_idx = batch_idx % mask_ne3;
 77        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
 78    }
 79
 80    ACC_TYPE4 q_priv[DK_VEC];
 81    if (my_query_row < n_q) {
 82        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
 83        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
 84        #pragma unroll
 85        for (int i = 0; i < DK_VEC; ++i) {
 86            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
 87        }
 88    }
 89
 90    ACC_TYPE4 o_acc[DV_VEC];
 91    #pragma unroll
 92    for (int i = 0; i < DV_VEC; ++i) {
 93        o_acc[i] = (ACC_TYPE4)(0.0f);
 94    }
 95    ACC_TYPE m_i = -INFINITY;
 96    ACC_TYPE l_i = 0.0f;
 97
 98    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
 99
100    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
101    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
102
103    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
104        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
105            const int row = i / DK_VEC;
106            const int col = i % DK_VEC;
107            const int k_row_idx = k_start + row;
108            if (k_row_idx < n_kv) {
109                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
110                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
111            }
112        }
113        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
114            const int row = i / DV_VEC;
115            const int col = i % DV_VEC;
116            const int v_row_idx = k_start + row;
117            if (v_row_idx < n_kv) {
118                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
119                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
120            }
121        }
122        barrier(CLK_LOCAL_MEM_FENCE);
123
124        if (my_query_row >= n_q) {
125            continue;
126        }
127
128        for (int j = 0; j < BLOCK_N; j += 2) {
129            const int k_row0 = k_start + j;
130            const int k_row1 = k_start + j + 1;
131
132            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
133            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
134            #pragma unroll
135            for (int k = 0; k < DK_VEC; k++) {
136                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
137                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
138            }
139            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
140            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
141
142            if (is_causal) {
143                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
144                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
145            }
146
147            if (k_row0 >= n_kv) score0 = -INFINITY;
148            if (k_row1 >= n_kv) score1 = -INFINITY;
149
150            if (mask_base != NULL) {
151                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
152                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
153                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
154            }
155
156            if (logit_softcap > 0.0f) {
157                score0 = logit_softcap * tanh(score0 / logit_softcap);
158                score1 = logit_softcap * tanh(score1 / logit_softcap);
159            }
160
161            const ACC_TYPE m_new = max(m_i, max(score0, score1));
162            const ACC_TYPE p0 = exp(score0 - m_new);
163            const ACC_TYPE p1 = exp(score1 - m_new);
164            const ACC_TYPE scale_prev = exp(m_i - m_new);
165
166            #pragma unroll
167            for (int i = 0; i < DV_VEC; ++i) {
168                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
169            }
170            l_i = l_i * scale_prev + p0 + p1;
171            m_i = m_new;
172        }
173    }
174
175    if (my_query_row < n_q) {
176        if (sinks_void != NULL) {
177            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
178            const ACC_TYPE m_sink = sinks_ptr[head_idx];
179            const ACC_TYPE m_final = max(m_i, m_sink);
180
181            const ACC_TYPE scale_o = exp(m_i - m_final);
182            #pragma unroll
183            for (int i = 0; i < DV_VEC; ++i) {
184                o_acc[i] *= scale_o;
185            }
186
187            l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final);
188        }
189
190        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
191        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
192        if (l_i > 0.0f) {
193            const ACC_TYPE l_inv = 1.0f / l_i;
194            #pragma unroll
195            for (int i = 0; i < DV_VEC; ++i) {
196                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
197            }
198        } else {
199            #pragma unroll
200            for (int i = 0; i < DV_VEC; ++i) {
201                o_row[i] = (DATA_TYPE4)(0.0f);
202            }
203        }
204    }
205}
206
207__kernel void flash_attn_f16_q1(
208    const global void * q_void, ulong q_offset,
209    const global void * k_void, ulong k_offset,
210    const global void * v_void, ulong v_offset,
211    global void * o_void, ulong o_offset,
212    const float scale,
213    const int n_q,
214    const int n_kv,
215    const int is_causal,
216    const int n_head,
217    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
218    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
219    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
220    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
221    const float max_bias,
222    const float m0,
223    const float m1,
224    const int n_head_log2,
225    const float logit_softcap,
226    const int n_head_kv,
227    const global void* mask_void,
228    const ulong mask_offset,
229    const ulong mask_nb1,
230    const ulong mask_nb2,
231    const ulong mask_nb3,
232    const int mask_ne2,
233    const int mask_ne3,
234    const global void* sinks_void,
235    const ulong sinks_offset
236) {
237    const int tid = get_local_id(0);
238    const int head_batch_idx = get_global_id(1);
239
240    const int batch_idx = head_batch_idx / n_head;
241    const int head_idx = head_batch_idx % n_head;
242
243    const int gqa_ratio = n_head / n_head_kv;
244    const int head_kv_idx = head_idx / gqa_ratio;
245
246    const global char* q_base = (const global char*)q_void + q_offset;
247    const global char* k_base = (const global char*)k_void + k_offset;
248    const global char* v_base = (const global char*)v_void + v_offset;
249    global char* o_base = (global char*)o_void + o_offset;
250
251    const global char* mask_base = NULL;
252    if (mask_void != NULL) {
253        const int mask_head_idx = head_idx % mask_ne2;
254        const int mask_batch_idx = batch_idx % mask_ne3;
255        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
256    }
257
258    ACC_TYPE4 q_priv[DK_VEC];
259    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
260    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
261    #pragma unroll
262    for (int i = 0; i < DK_VEC; ++i) {
263        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
264    }
265
266    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
267
268    const global ACC_TYPE* sinks_ptr = NULL;
269    if (sinks_void != NULL) {
270        sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
271    }
272
273    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
274    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
275        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
276        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
277        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
278        #pragma unroll
279        for (int k = 0; k < DK_VEC; k++) {
280            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
281        }
282        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
283        if (mask_base != NULL) {
284            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
285            score += slope * (ACC_TYPE)mask_ptr[k_idx];
286        }
287        if (logit_softcap > 0.0f) {
288            score = logit_softcap * tanh(score / logit_softcap);
289        }
290        m_i = max(m_i, score);
291    }
292
293    __local ACC_TYPE local_m[Q1_WG_SIZE];
294    local_m[tid] = m_i;
295    barrier(CLK_LOCAL_MEM_FENCE);
296    #pragma unroll
297    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
298        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
299        barrier(CLK_LOCAL_MEM_FENCE);
300    }
301    const ACC_TYPE m_final = local_m[0];
302
303    ACC_TYPE4 o_acc[DV_VEC];
304    #pragma unroll
305    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
306    ACC_TYPE l_i = 0.0f;
307
308    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
309        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
310        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
311        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
312        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
313        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
314        #pragma unroll
315        for (int k = 0; k < DK_VEC; k++) {
316            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
317        }
318        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
319        if (mask_base != NULL) {
320            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
321            score += slope * (ACC_TYPE)mask_ptr[k_idx];
322        }
323        if (logit_softcap > 0.0f) {
324            score = logit_softcap * tanh(score / logit_softcap);
325        }
326        const ACC_TYPE p = exp(score - m_final);
327        l_i += p;
328        #pragma unroll
329        for (int i = 0; i < DV_VEC; i++) {
330            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
331        }
332    }
333
334    __local ACC_TYPE local_l[Q1_WG_SIZE];
335    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
336    local_l[tid] = l_i;
337    barrier(CLK_LOCAL_MEM_FENCE);
338    #pragma unroll
339    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
340        if (tid < s) local_l[tid] += local_l[tid + s];
341        barrier(CLK_LOCAL_MEM_FENCE);
342    }
343
344    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
345    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
346    ACC_TYPE l_final = local_l[0];
347
348    if (sinks_ptr != NULL) {
349        l_final += exp(sinks_ptr[head_idx] - m_final);
350    }
351
352    if (l_final > 0.0f) {
353        const ACC_TYPE l_inv = 1.0f / l_final;
354        for (int i = 0; i < DV_VEC; i++) {
355            local_o_comp[tid] = o_acc[i];
356            barrier(CLK_LOCAL_MEM_FENCE);
357            #pragma unroll
358            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
359                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
360                barrier(CLK_LOCAL_MEM_FENCE);
361            }
362            if (tid == 0) {
363                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
364            }
365        }
366    } else if (tid == 0) {
367        #pragma unroll
368        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
369    }
370}