@@ -57,31 +57,33 @@ static __global__ void mul_mat_f(
5757 T * tile_xy = (T *) compute_base + threadIdx .y *(tile_A::I * tile_k_padded);
5858
5959 if constexpr (has_ids) {
60- __shared__ int has_any;
61- if (threadIdx .y == 0 ) {
62- int local_has_any = 0 ;
63- for (int j = threadIdx .x ; j < cols_per_block; j += warp_size) {
64- int slot = -1 ;
65- for (int k = 0 ; k < nchannels_dst; ++k) {
66- const int idv = ids[j*stride_row_id + k*stride_col_id];
67- if (idv == expert_idx) {
68- slot = k;
69- break ;
70- }
71- }
72- if (j < cols_per_block) {
73- local_has_any |= (slot >= 0 );
74- slot_map[j] = slot;
60+ int found = 0 ;
61+
62+ for (int j0 = 0 ; j0 < cols_per_block; j0 += nwarps) {
63+ const int j = j0 + threadIdx .y ;
64+ const int32_t * __restrict__ id_row = ids + j*stride_row_id;
65+
66+ if (threadIdx .x == 0 ) {
67+ slot_map[j] = -1 ;
68+ }
69+
70+ for (int k = threadIdx .x ; k < nchannels_dst; k += warp_size) {
71+ int match = id_row[k*stride_col_id] == expert_idx;
72+
73+ if (match) {
74+ slot_map[j] = k;
75+ found = 1 ;
76+ break ;
7577 }
7678 }
77- has_any = warp_reduce_any (local_has_any);
7879 }
79- __syncthreads ();
80- if (has_any == 0 ) {
80+
81+ if (! __syncthreads_or (found) ) {
8182 return ;
8283 }
8384 }
8485
86+
8587 for (int col = threadIdx .y *warp_size + threadIdx .x ; col < ncols; col += nwarps*warp_size) {
8688 tile_A A[ntA][warp_size / tile_A::J];
8789#pragma unroll
@@ -106,14 +108,7 @@ static __global__ void mul_mat_f(
106108 if constexpr (!has_ids) {
107109 tile_xy[j0*tile_k_padded + threadIdx .x ] = j < cols_per_block ? y[j*stride_col_y + col] : 0 .0f ;
108110 } else {
109- float val = 0 .0f ;
110- if (j < cols_per_block) {
111- const int slot = slot_map[j];
112- if (slot >= 0 ) {
113- val = y[slot*stride_channel_y + j*stride_col_y + col];
114- }
115- }
116- tile_xy[j0*tile_k_padded + threadIdx .x ] = val;
111+ tile_xy[j0*tile_k_padded + threadIdx .x ] = j < cols_per_block ? y[slot_map[j]*stride_channel_y + j*stride_col_y + col] : 0 .0f ;
117112 }
118113 }
119114 } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
@@ -125,14 +120,7 @@ static __global__ void mul_mat_f(
125120 const float2 tmp = j < cols_per_block ? y2[j*stride_col_y + col] : make_float2 (0 .0f , 0 .0f );
126121 tile_xy[j0*tile_k_padded + threadIdx .x ] = {tmp.x , tmp.y };
127122 } else {
128- float2 tmp = make_float2 (0 .0f , 0 .0f );
129- if (j < cols_per_block) {
130- const int slot = slot_map[j];
131- if (slot >= 0 ) {
132- const float2 * y2_slot = (const float2 *)(y + slot*stride_channel_y);
133- tmp = y2_slot[j*stride_col_y + col];
134- }
135- }
123+ float2 tmp = j < cols_per_block && slot_map[j] >= 0 ? *(const float2 *) &y[slot_map[j]*stride_channel_y + 2 *(j*stride_col_y + col)] : make_float2 (0 .0f , 0 .0f );
136124 tile_xy[j0*tile_k_padded + threadIdx .x ] = {tmp.x , tmp.y };
137125 }
138126 }
@@ -221,7 +209,7 @@ static inline void mul_mat_f_switch_ids(
221209 const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream) {
222210 if (ids) {
223211 mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true ><<<block_nums, block_dims, nbytes_shared_total, stream>>>
224- (x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
212+ (x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
225213 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
226214 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
227215 } else {
0 commit comments