before optimizing

src/operators/math/gemm_int8.cpp

@@ -60,19 +60,13 @@ void Gemm::AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
     "blt          1f                             \n\t"
     "0:                                          \n\t"
     "vld1.s8      {d0},       [%[a_ptr]],    r0  \n\t"  // A col0
-      "vld1.s8        {d1},       [%[a_ptr]],         r0  \n\t"  // A col1, q0
-      // used
-      "vld1.s8        {d2-d3},    [%[b_ptr]]!             \n\t"  // B row0, B
-      // row1, q1
-      // used
     "vmov.s8      q2,         #0                 \n\t"  // q2 used
+    "vld1.s8      {d1},       [%[a_ptr]],    r0  \n\t"  // A col1, q0 used
     "vdup.s8      d6,         d0[0]              \n\t"
+    "vld1.s8      {d2-d3},    [%[b_ptr]]!        \n\t"  // B row0, B row1, q1 used
     "vdup.s8      d7,         d1[0]              \n\t"  // q3 used
-      "vmlal.s8       q2,         d2,                 d6  \n\t"  // A col00 * B
-      // row0
-      "vmlal.s8       q2,         d3,                 d7  \n\t"  // A col10 * B
-      // row1, q3
-      // free
+    "vmlal.s8     q2,         d2,            d6  \n\t"  // A col00 * B row0
+    "vmlal.s8     q2,         d3,            d7  \n\t"  // A col10 * B row1, q3 free
     "vaddw.s16    q4,         q4,            d4  \n\t"
     "vaddw.s16    q5,         q5,            d5  \n\t"  // res row 0
     "vmov.s8      q2,                        #0  \n\t"
@@ -112,19 +106,13 @@ void Gemm::AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
     "vaddw.s16    q15,        q15,           d5  \n\t"  // res row 5
 
     "vld1.s8      {d0},       [%[a_ptr]],    r0  \n\t"  // A col0
-      "vld1.s8        {d1},       [%[a_ptr]],         r0  \n\t"  // A col1, q0
-      // used
-      "vld1.s8        {d2-d3},    [%[b_ptr]]!             \n\t"  // B row0, B
-      // row1, q1
-      // used
+    "vld1.s8      {d1},       [%[a_ptr]],    r0  \n\t"  // A col1, q0 used
+    "vld1.s8      {d2-d3},    [%[b_ptr]]!        \n\t"  // B row0, B row1, q1 used
     "vmov.s8      q2,         #0                 \n\t"  // q2 used
     "vdup.s8      d6,         d0[0]              \n\t"
     "vdup.s8      d7,         d1[0]              \n\t"  // q3 used
-      "vmlal.s8       q2,         d2,                 d6  \n\t"  // A col00 * B
-      // row0
-      "vmlal.s8       q2,         d3,                 d7  \n\t"  // A col10 * B
-      // row1, q3
-      // free
+    "vmlal.s8     q2,         d2,            d6  \n\t"  // A col00 * B row0
+    "vmlal.s8     q2,         d3,            d7  \n\t"  // A col10 * B row1, q3 free
     "vaddw.s16    q4,         q4,            d4  \n\t"
     "vaddw.s16    q5,         q5,            d5  \n\t"  // res row 0
     "vmov.s8      q2,                        #0  \n\t"
@@ -164,19 +152,13 @@ void Gemm::AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
     "vaddw.s16    q15,        q15,           d5  \n\t"  // res row 5
 
     "vld1.s8      {d0},       [%[a_ptr]],    r0  \n\t"  // A col0
-      "vld1.s8        {d1},       [%[a_ptr]],         r0  \n\t"  // A col1, q0
-      // used
-      "vld1.s8        {d2-d3},    [%[b_ptr]]!             \n\t"  // B row0, B
-      // row1, q1
-      // used
+    "vld1.s8      {d1},       [%[a_ptr]],    r0  \n\t"  // A col1, q0 used
+    "vld1.s8      {d2-d3},    [%[b_ptr]]!        \n\t"  // B row0, B row1, q1 used
     "vmov.s8      q2,         #0                 \n\t"  // q2 used
     "vdup.s8      d6,         d0[0]              \n\t"
     "vdup.s8      d7,         d1[0]              \n\t"  // q3 used
-      "vmlal.s8       q2,         d2,                 d6  \n\t"  // A col00 * B
-      // row0
-      "vmlal.s8       q2,         d3,                 d7  \n\t"  // A col10 * B
-      // row1, q3
-      // free
+    "vmlal.s8     q2,         d2,            d6  \n\t"  // A col00 * B row0
+    "vmlal.s8     q2,         d3,            d7  \n\t"  // A col10 * B row1, q3 free
     "vaddw.s16    q4,         q4,            d4  \n\t"
     "vaddw.s16    q5,         q5,            d5  \n\t"  // res row 0
     "vmov.s8      q2,                        #0  \n\t"
@@ -216,19 +198,13 @@ void Gemm::AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
     "vaddw.s16    q15,        q15,           d5  \n\t"  // res row 5
 
     "vld1.s8      {d0},       [%[a_ptr]],    r0  \n\t"  // A col0
-      "vld1.s8        {d1},       [%[a_ptr]],         r0  \n\t"  // A col1, q0
-      // used
-      "vld1.s8        {d2-d3},    [%[b_ptr]]!             \n\t"  // B row0, B
-      // row1, q1
-      // used
+    "vld1.s8      {d1},       [%[a_ptr]],    r0  \n\t"  // A col1, q0 used
+    "vld1.s8      {d2-d3},    [%[b_ptr]]!        \n\t"  // B row0, B row1, q1 used
     "vmov.s8      q2,         #0                 \n\t"  // q2 used
     "vdup.s8      d6,         d0[0]              \n\t"
     "vdup.s8      d7,         d1[0]              \n\t"  // q3 used
-      "vmlal.s8       q2,         d2,                 d6  \n\t"  // A col00 * B
-      // row0
-      "vmlal.s8       q2,         d3,                 d7  \n\t"  // A col10 * B
-      // row1, q3
-      // free
+    "vmlal.s8     q2,         d2,            d6  \n\t"  // A col00 * B row0
+    "vmlal.s8     q2,         d3,            d7  \n\t"  // A col10 * B row1, q3 free
     "vaddw.s16    q4,         q4,            d4  \n\t"
     "vaddw.s16    q5,         q5,            d5  \n\t"  // res row 0
     "vmov.s8      q2,                        #0  \n\t"
@@ -274,19 +250,13 @@ void Gemm::AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
     "blt          3f                             \n\t"
     "2:                                          \n\t"
     "vld1.s8      {d0},       [%[a_ptr]],    r0  \n\t"  // A col0
-      "vld1.s8        {d1},       [%[a_ptr]],         r0  \n\t"  // A col1, q0
-      // used
-      "vld1.s8        {d2-d3},    [%[b_ptr]]!             \n\t"  // B row0, B
-      // row1, q1
-      // used
+    "vld1.s8      {d1},       [%[a_ptr]],    r0  \n\t"  // A col1, q0 used
+    "vld1.s8      {d2-d3},    [%[b_ptr]]!        \n\t"  // B row0, B row1, q1 used
     "vmov.s8      q2,         #0                 \n\t"  // q2 used
     "vdup.s8      d6,         d0[0]              \n\t"
     "vdup.s8      d7,         d1[0]              \n\t"  // q3 used
-      "vmlal.s8       q2,         d2,                 d6  \n\t"  // A col00 * B
-      // row0
-      "vmlal.s8       q2,         d3,                 d7  \n\t"  // A col10 * B
-      // row1, q3
-      // free
+    "vmlal.s8     q2,         d2,            d6  \n\t"  // A col00 * B row0
+    "vmlal.s8     q2,         d3,            d7  \n\t"  // A col10 * B row1, q3 free
     "vaddw.s16    q4,         q4,            d4  \n\t"
     "vaddw.s16    q5,         q5,            d5  \n\t"  // res row 0
     "vmov.s8      q2,                        #0  \n\t"
@@ -323,13 +293,12 @@ void Gemm::AddDot6x8(int32_t k, const int8_t *a, const int8_t *b, int32_t *c,
     "vmlal.s8     q2,         d2,            d6  \n\t"
     "vmlal.s8     q2,         d3,            d7  \n\t"
     "vaddw.s16    q14,        q14,           d4  \n\t"
-      "vaddw.s16      q15,        q15,                d5  \n\t"  // res row 5
+    "vaddw.s16    q15,        q15,           d5  \n\t"  // res row5
 
     "subs         %[kc3],     %[kc3],        #1  \n\t"
     "bge          2b                             \n\t"
 
-      "3:                                                 \n\t"  // odd, last
-      // row
+    "3:                                          \n\t"  // odd, last row
     "subs         %[kc4],     %[kc4],        #1  \n\t"
     "blt          4f                             \n\t"
     "vld1.s8      {d0},       [%[a_ptr]]         \n\t"

test/operators/test_mul_op.cpp

@@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include <stdint-gcc.h>
 #include "../test_helper.h"
 #include "../test_include.h"
 #include "operators/mul_op.h"
@@ -73,12 +74,19 @@ int TestMulOP() {
     }
   }
 
+  int32_t eq = 0;
+  int32_t neq = 0;
   for (int32_t i = 0; i < m * n; ++i) {
     PADDLE_MOBILE_ENFORCE(
         output_data[i] == c[i], "output[%d] = %d, output_cmp[%d] = %d", i,
         static_cast<int32_t>(output_data[i]), i, static_cast<int32_t>(c[i]));
+    if (static_cast<int>(output_data[i] == c[i])) {
+      ++eq;
+    } else {
+      ++neq;
     }
-  DLOG << "Run MulOp successfully!";
+  }
+  DLOG << "mnk=" << m << " " << n << " " << k << "   eq=" << eq << " neq=" << neq;
   delete op;
   return 0;
 }