Merged
docs: 修复注释与代码不一致(asinh dtype 范围与 cdist UB 注释) #6492
docs: 修复注释与代码不一致(asinh dtype 范围与 cdist UB 注释) #6492
Merged
TangPCAdd on 19 days ago
Total: 2 files changed+5-4
@@ -19,8 +19,8 @@
19 * - PrecisionReduceFlag(false):19 步内部已升 FP32 精度,无需框架降精度19 * - PrecisionReduceFlag(false):19 步内部已升 FP32 精度,无需框架降精度
20 * - ascend950、ascend350(DAV_3510 / arch35,复用 arch35 tiling)20 * - ascend950、ascend350(DAV_3510 / arch35,复用 arch35 tiling)
21 *21 *
22- * 迭代一范围:仅 FP32 完整实现;FP16/BF16 在迭代二落地。22+ * dtype 范围:FP32、FP16、BF16 均已实现。
23- * 但 DataType 列表声明 3 个 dtype 保证 TilingKey 接口稳定。23+ * DataType 列表声明 3 个 dtype 保证 TilingKey 接口稳定。
24 */24 */
25#include "register/op_def_registry.h"25#include "register/op_def_registry.h"
26 26 
@@ -485,8 +485,9 @@ void CdistTiling::DoTiling()
485 if (M_ >= kHlMMin && M_ <= M_SIZE && hlVectorWins) {485 if (M_ >= kHlMMin && M_ <= M_SIZE && hlVectorWins) {
486 is_small_m_ = 0; // launch on the Normal-shaped tiling486 is_small_m_ = 0; // launch on the Normal-shaped tiling
487 use_reduce_hl_ = 1; // but dispatch to the high-level reduce kernel (KERNEL_MODE=3)487 use_reduce_hl_ = 1; // but dispatch to the high-level reduce kernel (KERNEL_MODE=3)
488- // Reserve UB for the two fixed HL fp32 compute planes (x1exp + diff, 4096 elems each = 32KB) plus488+ // Reserve UB for the two fixed HL fp32 compute planes (x1exp + diff, 8192 elems each = 32KB per
489- // the fp32 M-split accumulator margin, so the extra planes never overflow UB on large P·R tiles.489+ // plane, 64KB total) plus the fp32 M-split accumulator margin, so the extra planes never overflow
490+ // UB on large P·R tiles.
490 int64_t savedUb = ubSize_;491 int64_t savedUb = ubSize_;
491 ubSize_ = ubSize_ - HL_UB_RESERVE;492 ubSize_ = ubSize_ - HL_UB_RESERVE;
492 if (ubSize_ < BLOCK_BYTES)493 if (ubSize_ < BLOCK_BYTES)