@@ -124,27 +124,27 @@ TEST_P(Execution, Dma_01_Direct) {
124
124
}
125
125
126
126
TEST_P (Execution, Dma_02_Explicit_Copy) {
127
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
128
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
127
+ AddLocalBuffer<cl_int> (local_wg_size);
128
+ AddLocalBuffer<cl_int> (local_wg_size);
129
129
AddInputBuffer (kts::N, vaddInA);
130
130
AddInputBuffer (kts::N, vaddInB);
131
131
AddOutputBuffer (kts::N, vaddOutC);
132
132
RunGeneric1D (kts::N, local_wg_size);
133
133
}
134
134
135
135
TEST_P (Execution, Dma_03_Explicit_Copy_Rotate) {
136
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
137
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
136
+ AddLocalBuffer<cl_int> (local_wg_size);
137
+ AddLocalBuffer<cl_int> (local_wg_size);
138
138
AddInputBuffer (kts::N, vaddInA);
139
139
AddInputBuffer (kts::N, vaddInB);
140
140
AddOutputBuffer (kts::N, vaddOutC);
141
141
RunGeneric1D (kts::N, local_wg_size);
142
142
}
143
143
144
144
TEST_P (Execution, Dma_04_async_copy) {
145
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
146
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
147
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
145
+ AddLocalBuffer<cl_int> (local_wg_size);
146
+ AddLocalBuffer<cl_int> (local_wg_size);
147
+ AddLocalBuffer<cl_int> (local_wg_size);
148
148
AddInputBuffer (kts::N, vaddInA);
149
149
AddInputBuffer (kts::N, vaddInB);
150
150
AddOutputBuffer (kts::N, vaddOutC);
@@ -156,12 +156,12 @@ TEST_P(Execution, Dma_04_async_copy) {
156
156
// size of the local buffers.
157
157
TEST_P (Execution, Dma_05_async_double_buffer) {
158
158
const cl_int iterations = 16 ;
159
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
160
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
161
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
162
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
163
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
164
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
159
+ AddLocalBuffer<cl_int> (local_wg_size);
160
+ AddLocalBuffer<cl_int> (local_wg_size);
161
+ AddLocalBuffer<cl_int> (local_wg_size);
162
+ AddLocalBuffer<cl_int> (local_wg_size);
163
+ AddLocalBuffer<cl_int> (local_wg_size);
164
+ AddLocalBuffer<cl_int> (local_wg_size);
165
165
AddInputBuffer (kts::N * iterations, vaddInA);
166
166
AddInputBuffer (kts::N * iterations, vaddInB);
167
167
AddOutputBuffer (kts::N * iterations, vaddOutC);
@@ -267,10 +267,9 @@ TEST_P(AsyncCopyTests, Dma_10_half_async_copy) {
267
267
const auto param = getParam ();
268
268
AddMacro (" TYPE" , param.type_str );
269
269
270
- const size_t local_buffer_len = local_wg_size * param.type_size ;
271
- AddLocalBuffer (local_buffer_len);
272
- AddLocalBuffer (local_buffer_len);
273
- AddLocalBuffer (local_buffer_len);
270
+ AddLocalBuffer (local_wg_size, param.type_size );
271
+ AddLocalBuffer (local_wg_size, param.type_size );
272
+ AddLocalBuffer (local_wg_size, param.type_size );
274
273
275
274
if (3 == param.vec_width ) {
276
275
AddInputBuffer (kts::N, makeHalf3Streamer (HalfTypeParam::InA));
@@ -299,10 +298,9 @@ TEST_P(AsyncCopyTests, Dma_11_half_async_strided_copy) {
299
298
const auto param = getParam ();
300
299
AddMacro (" TYPE" , param.type_str );
301
300
302
- const size_t local_buffer_len = local_wg_size * param.type_size ;
303
- AddLocalBuffer (local_buffer_len);
304
- AddLocalBuffer (local_buffer_len);
305
- AddLocalBuffer (local_buffer_len);
301
+ AddLocalBuffer (local_wg_size, param.type_size );
302
+ AddLocalBuffer (local_wg_size, param.type_size );
303
+ AddLocalBuffer (local_wg_size, param.type_size );
306
304
307
305
if (3 == param.vec_width ) {
308
306
AddInputBuffer (kts::N * 2 , makeHalf3Streamer (HalfTypeParam::InA));
@@ -376,8 +374,8 @@ TEST_P(Execution, Dma_13_wait_event_is_barrier) {
376
374
(((x % local_wg_size) + 1 ) % local_wg_size));
377
375
};
378
376
379
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
380
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
377
+ AddLocalBuffer<cl_int> (local_wg_size);
378
+ AddLocalBuffer<cl_int> (local_wg_size);
381
379
AddInputBuffer (kts::N, vaddInA);
382
380
AddInputBuffer (kts::N, vaddInB);
383
381
AddOutputBuffer (kts::N, rotateB);
@@ -389,7 +387,7 @@ TEST_P(Execution, Dma_14_wait_event_is_barrier_overwrite) {
389
387
return vaddInA (x) + 1 ;
390
388
};
391
389
392
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
390
+ AddLocalBuffer<cl_int> (local_wg_size);
393
391
AddInputBuffer (kts::N, vaddInA);
394
392
AddOutputBuffer (kts::N, vaddInAPlusOne);
395
393
RunGeneric1D (kts::N, local_wg_size);
@@ -408,10 +406,10 @@ TEST_P(Execution, DISABLED_Dma_15_wait_event_is_execution_barrier) {
408
406
(((x % local_wg_size) + 1 ) % local_wg_size));
409
407
};
410
408
411
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
412
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
413
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
414
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
409
+ AddLocalBuffer<cl_int> (local_wg_size);
410
+ AddLocalBuffer<cl_int> (local_wg_size);
411
+ AddLocalBuffer<cl_int> (local_wg_size);
412
+ AddLocalBuffer<cl_int> (local_wg_size);
415
413
AddInputBuffer (kts::N, vaddInA);
416
414
AddInputBuffer (kts::N, vaddInB);
417
415
AddOutputBuffer (kts::N, rotateA);
@@ -424,7 +422,7 @@ TEST_P(Execution, Dma_16_wait_event_is_barrier_strided) {
424
422
return vaddInA (x) + 1 ;
425
423
};
426
424
427
- AddLocalBuffer (local_wg_size * sizeof (cl_int) );
425
+ AddLocalBuffer<cl_int> (local_wg_size);
428
426
AddInputBuffer (kts::N, vaddInA);
429
427
AddOutputBuffer (kts::N, vaddInAPlusOne);
430
428
RunGeneric1D (kts::N, local_wg_size);
0 commit comments