Merged
fix(max_pool3d_grad): 修复 NDHWC C 轴切分 UB 越界 #13160
小王!Add on 22 days ago
fix(max_pool3d_grad): 修复 NDHWC C 轴切分 UB 越界 #13160
Merged
Total: 2 files changed+7-7
| @@ -620,8 +620,8 @@ TEST_F(MaxPool3DGradNdhwcBigTiling, non_overlap_k_eq_s_claimed_by_big_kernel) | |||
| 620 | gert::StorageShape dxShape = {{2, 13, 13, 65, 8}, {2, 13, 13, 65, 8}}; | 620 | gert::StorageShape dxShape = {{2, 13, 13, 65, 8}, {2, 13, 13, 65, 8}}; |
| 621 | auto attrs = MakeAttrs({1, 4, 4, 8, 1}, {1, 4, 4, 8, 1}, "VALID", {0, 0, 0, 0, 0, 0}); | 621 | auto attrs = MakeAttrs({1, 4, 4, 8, 1}, {1, 4, 4, 8, 1}, "VALID", {0, 0, 0, 0, 0, 0}); |
| 622 | RunBigTilingCase(xShape, yShape, dxShape, attrs, {1, 4, 4, 8, 1}, {1, 4, 4, 8, 1}, ge::DT_FLOAT16, | 622 | RunBigTilingCase(xShape, yShape, dxShape, attrs, {1, 4, 4, 8, 1}, {1, 4, 4, 8, 1}, ge::DT_FLOAT16, |
| 623 | - GET_TPL_TILING_KEY(TPL_INT32, 0, 1, 0, 0), true, 2048, 288, 64, 8, 16384, 1024, 32768, 4, 4, 4, 4, | 623 | + GET_TPL_TILING_KEY(TPL_INT32, 0, 1, 0, 0), true, 256, 288, 8, 8, 4096, 256, 8192, 4, 4, 4, 4, 9, |
| 624 | - 9, 8); | 624 | + 8); |
| 625 | } | 625 | } |
| 626 | 626 | ||
| 627 | // k < s(g1_11/g1_12 形态): kv=128, C=8, fp16 —— 应被大 kernel 认领 | 627 | // k < s(g1_11/g1_12 形态): kv=128, C=8, fp16 —— 应被大 kernel 认领 |
| @@ -633,8 +633,8 @@ TEST_F(MaxPool3DGradNdhwcBigTiling, non_overlap_k_lt_s_claimed_by_big_kernel) | |||
| 633 | gert::StorageShape dxShape = {{2, 13, 13, 65, 8}, {2, 13, 13, 65, 8}}; | 633 | gert::StorageShape dxShape = {{2, 13, 13, 65, 8}, {2, 13, 13, 65, 8}}; |
| 634 | auto attrs = MakeAttrs({1, 4, 4, 8, 1}, {1, 7, 6, 11, 1}, "VALID", {0, 0, 0, 0, 0, 0}); | 634 | auto attrs = MakeAttrs({1, 4, 4, 8, 1}, {1, 7, 6, 11, 1}, "VALID", {0, 0, 0, 0, 0, 0}); |
| 635 | RunBigTilingCase(xShape, yShape, dxShape, attrs, {1, 4, 4, 8, 1}, {1, 7, 6, 11, 1}, ge::DT_FLOAT16, | 635 | RunBigTilingCase(xShape, yShape, dxShape, attrs, {1, 4, 4, 8, 1}, {1, 7, 6, 11, 1}, ge::DT_FLOAT16, |
| 636 | - GET_TPL_TILING_KEY(TPL_INT32, 0, 1, 0, 0), true, 2048, 72, 64, 8, 16384, 1024, 118272, 2, 7, 3, 6, | 636 | + GET_TPL_TILING_KEY(TPL_INT32, 0, 1, 0, 0), true, 256, 72, 8, 8, 4096, 256, 29568, 2, 7, 3, 6, 6, |
| 637 | - 6, 11); | 637 | + 11); |
| 638 | } | 638 | } |
| 639 | 639 | ||
| 640 | // k == s + 大 C(g1_10 精确形态): kv=128, C=64, fp16 —— 大 kernel 认领, | 640 | // k == s + 大 C(g1_10 精确形态): kv=128, C=64, fp16 —— 大 kernel 认领, |
| @@ -208,18 +208,18 @@ bool Pool3DGradNDHWCSmallKernelCommonTiling::TrySplitAlignC() | |||
| 208 | if (IsMeetUBSize() && IsMeetTargetCoreNum()) { | 208 | if (IsMeetUBSize() && IsMeetTargetCoreNum()) { |
| 209 | int64_t left = 1; | 209 | int64_t left = 1; |
| 210 | int64_t right = Ops::Base::CeilDiv(inputData->cX / 2, baseData.moveDataNumCacheLine); | 210 | int64_t right = Ops::Base::CeilDiv(inputData->cX / 2, baseData.moveDataNumCacheLine); |
| 211 | - int64_t bestSplit = 1; | 211 | + int64_t bestInner = tmpCAligned; |
| 212 | while (left <= right) { | 212 | while (left <= right) { |
| 213 | int64_t mid = left + (right - left) / 2; | 213 | int64_t mid = left + (right - left) / 2; |
| 214 | splitData.cOutputInner = mid * baseData.moveDataNumCacheLine; | 214 | splitData.cOutputInner = mid * baseData.moveDataNumCacheLine; |
| 215 | if (IsMeetUBSize() && IsMeetTargetCoreNum()) { | 215 | if (IsMeetUBSize() && IsMeetTargetCoreNum()) { |
| 216 | - bestSplit = mid; | 216 | + bestInner = splitData.cOutputInner; |
| 217 | left = mid + 1; | 217 | left = mid + 1; |
| 218 | } else { | 218 | } else { |
| 219 | right = mid - 1; | 219 | right = mid - 1; |
| 220 | } | 220 | } |
| 221 | } | 221 | } |
| 222 | - splitData.cOutputInner = bestSplit * baseData.moveDataNumCacheLine; | 222 | + splitData.cOutputInner = std::min(bestInner, inputData->cX); |
| 223 | if (splitData.cOutputInner > baseData.proDataNumInOneBeat) { | 223 | if (splitData.cOutputInner > baseData.proDataNumInOneBeat) { |
| 224 | splitData.cOutputInner = baseData.proDataNumInOneBeat; | 224 | splitData.cOutputInner = baseData.proDataNumInOneBeat; |
| 225 | } | 225 | } |