comments

Signed-off-by: Liqun Fu <[email protected]>

onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx_common_int8.h

@@ -20,7 +20,6 @@ SQ4BitGemmM1Kernel_CompInt8_avx2(
     const float* Bias
 );
 
-#include <iostream>
 template <size_t NCols, bool HasZeroPoint>
 MLAS_FORCEINLINE void
 ComputeDotProducts_BlkBitWidth4_CompInt8_SubBlkLen16(
@@ -575,6 +574,7 @@ ComputeDotProducts_BlkBitWidth4_CompInt8_SubBlkLen64_NCols4(
     _mm_storeu_ps(SumPtr, acc_x);
 }
 
+// TODO: is this able to be inlined if DotQuadFunctionType is a function pointer?
 template <bool HasZeroPoint, DotQuadFunctionType dot_quad>
 MLAS_FORCEINLINE void
 ComputeDotProducts_BlkBitWidth4_CompInt8_SubBlkLen64_NCols1(
@@ -622,6 +622,8 @@ ComputeDotProducts_BlkBitWidth4_CompInt8_SubBlkLen64_NCols1(
         uint8_t zp0;
         if constexpr (HasZeroPoint) {
             // TODO: this block causes near 30% of the computation.
+            // The solution proposed by @yufenglee is to factor out scaleB * zp
+            // while packing A. Will do this in next PR.
             bool is_lower = (QuantBZeroPointIdx & 1) == 0;
             std::byte zp_packed = QuantBZeroPointColPtr[0 * StrideQuantBZeroPoint + QuantBZeroPointIdx / 2];
             zp0 = std::to_integer<int8_t>(is_lower ? (zp_packed & std::byte{0x0F}) : (zp_packed >> 4));