From 3182f22aa7114dc45f2621d45cffa3eb6100de16 Mon Sep 17 00:00:00 2001
From: YeZZ <yezhenni1@huawei.com>
Date: Thu, 15 May 2025 09:46:44 +0800
Subject: [PATCH] A fullload constant tiling

---
 .../tiling/matmul_constant_tiling_impl.h      | 40 +++++++++++++++++++
 .../tiling/matmul_constant_tiling_utils.h     |  3 ++
 lib/matmul/constant_tiling.h                  | 14 ++++---
 3 files changed, 51 insertions(+), 6 deletions(-)

diff --git a/impl/matmul/tiling/matmul_constant_tiling_impl.h b/impl/matmul/tiling/matmul_constant_tiling_impl.h
index b1691c9d..1d90462a 100644
--- a/impl/matmul/tiling/matmul_constant_tiling_impl.h
+++ b/impl/matmul/tiling/matmul_constant_tiling_impl.h
@@ -429,5 +429,45 @@ __aicore__ constexpr L1Status GetL1Factor(const MatmulConfig &mmCFG, int32_t l1S
     }
     return neitherFullLoad;
 }
+
+template <typename A_TYPE, typename B_TYPE, typename C_TYPE, typename BIAS_TYPE>
+__aicore__ constexpr bool CalcAL1FullLoadTiling(int32_t l1Size, MatmulApiStaticTiling &tiling)
+{
+    if (tiling.singleCoreM > Impl::MIN_MN_SIZE || tiling.singleCoreN > Impl::MIN_MN_SIZE) {
+        return false;
+    }
+    l1Size = tiling.isBias ? (l1Size - Impl::BT_SIZE / Impl::BITS_PER_BYTE) : l1Size;
+    int32_t baseKaAlign = Align<int32_t>(tiling.baseK,
+        GetKAAlignValue<A_TYPE>() * GetReduceC0Size<typename A_TYPE::T>());
+    int32_t baseKbAlign = Align<int32_t>(tiling.baseK,
+        GetKBAlignValue<A_TYPE, B_TYPE>() * GetReduceC0Size<typename B_TYPE::T>());
+    int32_t baseA = tiling.baseM * baseKaAlign * GetBitSize<typename A_TYPE::T>() / Impl::BITS_PER_BYTE;
+    int32_t baseB = tiling.baseN * baseKbAlign * GetBitSize<typename B_TYPE::T>() / Impl::BITS_PER_BYTE;
+    int32_t depthA1 = (l1Size - Impl::DB_ON * baseB) / baseA;
+    if (depthA1 * tiling.baseM * baseKaAlign < tiling.singleCoreM * tiling.singleCoreK) {
+        return false;
+    }
+    depthA1 = MaxValue(tiling.singleCoreM, tiling.baseM) * MaxValue(tiling.singleCoreK, baseKaAlign) /
+        tiling.baseM / baseKaAlign;
+    tiling.depthA1 = depthA1;
+    tiling.stepKa = depthA1;
+
+    int32_t stepKb = (l1Size - depthA1 * baseA) / baseB / Impl::DB_ON;
+    if (stepKb * Impl::DB_ON * baseB > tiling.singleCoreK * tiling.singleCoreN) {
+        stepKb = MaxValue(tiling.singleCoreK, baseKbAlign) * MaxValue(tiling.singleCoreN, tiling.baseN) /
+            tiling.baseN / baseKbAlign / Impl::DB_ON;
+    }
+    if (stepKb < 1) {
+        tiling.depthB1 = 1;
+        tiling.stepKb = 1;
+        return true;
+    }
+    while (tiling.stepKa % stepKb != 0 && stepKb % tiling.stepKa != 0 && stepKb > 1) {
+        stepKb--;
+    }
+    tiling.depthB1 = stepKb * Impl::DB_ON;
+    tiling.stepKb = stepKb;
+    return true;
+}
 } // namespace AscendC
 #endif // _MATMUL_CONSTANT_TILING_IMPL_
\ No newline at end of file
diff --git a/impl/matmul/tiling/matmul_constant_tiling_utils.h b/impl/matmul/tiling/matmul_constant_tiling_utils.h
index e9c67d01..01166a86 100644
--- a/impl/matmul/tiling/matmul_constant_tiling_utils.h
+++ b/impl/matmul/tiling/matmul_constant_tiling_utils.h
@@ -25,6 +25,9 @@ constexpr int32_t DB_ON = 2;
 constexpr int32_t DB_OFF = 1;
 constexpr int32_t MIN_MTE1_LOAD = 32;
 constexpr int32_t OUTER_STEP = 2;
+constexpr int32_t BT_SIZE = 1024;
+constexpr int32_t MIN_MN_SIZE = 16;
+constexpr int32_t BITS_PER_BYTE = 8;
 #if __CCE_AICORE__ < 220
 constexpr int32_t L1_SIZE = 1024 * 1024;
 #elif __CCE_AICORE__ == 300
diff --git a/lib/matmul/constant_tiling.h b/lib/matmul/constant_tiling.h
index 7d589f7e..63eaa5df 100644
--- a/lib/matmul/constant_tiling.h
+++ b/lib/matmul/constant_tiling.h
@@ -46,16 +46,18 @@ __aicore__ constexpr MatmulApiStaticTiling GetMatmulApiTiling(const MatmulConfig
     tiling.baseM = mmCFG.basicM;
     tiling.baseN = mmCFG.basicN;
     tiling.baseK = mmCFG.basicK;
+    tiling.isBias = mmCFG.enableSetBias;
     tiling.stepM = l1Factor.mAL1;
     tiling.stepN = l1Factor.nBL1;
     int32_t reduceC0Size = GetReduceC0Size<typename A_TYPE::T>();
-    int32_t kL0 = GetKL0<A_TYPE>(mmCFG);
-    tiling.stepKa = CeilNoLog<int32_t>(l1Factor.kAL1, kL0);
-    tiling.stepKb = CeilNoLog<int32_t>(l1Factor.kBL1, kL0);
-    tiling.depthA1 = CeilNoLog<int32_t>(l1Factor.kAL1, kL0) * l1Factor.mAL1 * l1Factor.dbAL1;
-    tiling.depthB1 = CeilNoLog<int32_t>(l1Factor.kBL1, kL0) * l1Factor.nBL1 * l1Factor.dbBL1;
+    if (!CalcAL1FullLoadTiling<A_TYPE, B_TYPE, C_TYPE, BIAS_TYPE>(l1Size, tiling)) {
+        int32_t kL0 = GetKL0<A_TYPE>(mmCFG);
+        tiling.stepKa = CeilNoLog<int32_t>(l1Factor.kAL1, kL0);
+        tiling.stepKb = CeilNoLog<int32_t>(l1Factor.kBL1, kL0);
+        tiling.depthA1 = CeilNoLog<int32_t>(l1Factor.kAL1, kL0) * l1Factor.mAL1 * l1Factor.dbAL1;
+        tiling.depthB1 = CeilNoLog<int32_t>(l1Factor.kBL1, kL0) * l1Factor.nBL1 * l1Factor.dbBL1;
+    }
     tiling.iterateOrder = GetIterateOrder<A_TYPE, B_TYPE, C_TYPE, BIAS_TYPE>(l1Factor, mmCFG);
-    tiling.isBias = mmCFG.enableSetBias;
     tiling.dbL0A = GetL0ADb(mmCFG, TOTAL_L0A_SIZE);
     tiling.dbL0B = GetL0BDb(mmCFG, TOTAL_L0B_SIZE);
     // keep the same with runtime tiling, fix l0c db
-- 
Gitee