*fix bugs: Error in Base implementation, AMX-BF16 optimizations of class SynetInnerProduct16bGemmNN (part 2).

Author: ermig1979

src/Simd/SimdAmxBf16SynetInnerProduct16bGemmNN.cpp

@@ -32,6 +32,7 @@
 #include "Simd/SimdBFloat16.h"
 #include "Simd/SimdCopy.h"
 #include "Simd/SimdTile.h"
+#include "Simd/SimdLog.h"
 
 namespace Simd
 {
@@ -291,7 +292,7 @@ namespace Simd
             size_t M, size_t N, size_t K, int update, const uint16_t* B, float* C, int post, const float* bias, uint8_t* dst)
         {
             size_t m = 32, m1 = M, mm = AlignLo(m1, m), t = m1 - mm;
-            size_t dA = a.aK, dB = a.bK * DF, dC = (a.macroK < a.aK || a.macroN != a.aN || a.macroM != a.aM) ? a.cN : 0, dD = p.N * a.eC;
+            size_t dA = a.aK, dB = a.bK * DF, dC = (a.macroK < a.aK || a.macroN != a.aN || a.macroM != a.aM || C == (float*)dst) ? a.cN : 0, dD = p.N * a.eC;
             __m512 _bias[2];
             if (mm)
             {

src/Simd/SimdBaseSynetInnerProduct16bGemmNN.cpp

@@ -129,7 +129,7 @@ namespace Simd
             a.aM = AlignHi(p.M, a.microM);
             a.macroK = Simd::RestrictRange(AlignLo(L1 / a.microN / 2, a.microK), a.microK, a.aK);
             a.macroN = Simd::RestrictRange(AlignLo(L3 / a.macroK / 2, a.microN), a.microN, a.aN);
-            a.macroM = Simd::RestrictRange(L2 / a.macroK / 2, a.microM, a.aM);
+            a.macroM = Simd::RestrictRange(AlignLo(L2 / a.macroK / 2, a.microM), a.microM, a.aM);
             a.eA = p.typeA == SimdTensorData32f ? 4 : 2;
             a.eB = p.typeB == SimdTensorData32f ? 4 : 2;
             a.eC = p.typeC == SimdTensorData32f ? 4 : 2;

src/Test/TestSynetInnerProduct16b.cpp

@@ -198,15 +198,16 @@ namespace Test
         result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(128, 127, 128, b16, b16, f32, f, f, t), f1, f2);
 #endif
 #if 1
+        result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(64, 512, 512, f32, f32, f32, f, t, t), f1, f2);
+        result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(64, 608, 608, f32, f32, f32, f, t, t), f1, f2);
         result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(1824, 64, 608, f32, f32, f32, f, t, t), f1, f2);
         result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(64, 1824, 608, f32, f32, f32, f, t, t), f1, f2);
 #endif
 #if 0
         result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(3333, 3333, 3333, b16, b16, f32, f, f, t), f1, f2);
 #endif
 #else
-        result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(1824, 64, 608, f32, f32, f32, f, t, t), f1, f2);
-        result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(64, 1824, 608, f32, f32, f32, f, t, t), f1, f2);
+        result = result && SynetInnerProduct16bForwardAutoTest(eps, Param(64, 512, 512, f32, f32, f32, f, t, t), f1, f2);
 #endif
 
         return result;