[BesTLA] AVX2: Use loaded registers of B. (#151)

parvizmp · web-flow · commit aa4a8abdf8f9 · 2024-03-05T13:26:11.000+08:00
diff --git a/bestla/bestla/bestla_gemm.h b/bestla/bestla/bestla_gemm.h
@@ -2716,7 +2716,7 @@ class AvxvnniN8P4 : protected bestla::xbyak::JitAvxvnni {
           vpbroadcastd(vreg_t(AReg), ptr[reg_tmp1]);
           add(reg_tmp1, reg_astride);
           for (int i = 0; i < NRegs; i++) {
-            vpdpbusds_(vreg_t(CReg + mm * NRegs + i), vreg_t(AReg), ptr[reg_matBptr + kk * BKStepSize + i * VecBytes]);
+            vpdpbusds_(vreg_t(CReg + mm * NRegs + i), vreg_t(AReg), vreg_t(BReg + i));
           }
         }
       }

Original file line number	Diff line number	Diff line change
`@@ -2716,7 +2716,7 @@ class AvxvnniN8P4 : protected bestla::xbyak::JitAvxvnni {`
`2716`	`2716`	`vpbroadcastd(vreg_t(AReg), ptr[reg_tmp1]);`
`2717`	`2717`	`add(reg_tmp1, reg_astride);`
`2718`	`2718`	`for (int i = 0; i < NRegs; i++) {`
`2719`		`- vpdpbusds_(vreg_t(CReg + mm * NRegs + i), vreg_t(AReg), ptr[reg_matBptr + kk * BKStepSize + i * VecBytes]);`
	`2719`	`+ vpdpbusds_(vreg_t(CReg + mm * NRegs + i), vreg_t(AReg), vreg_t(BReg + i));`
`2720`	`2720`	`}`
`2721`	`2721`	`}`
`2722`	`2722`	`}`