!5182 optimization for winograd matmul

Merge pull request !5182 from lixian/master

mindspore/lite/nnacl/assembly/arm64/MatmulFp32OptRemain.S

@@ -0,0 +1,144 @@
+#ifdef __aarch64__
+    .text
+    .align 5
+    .global MatmulFloatNeon64OptRemain
+#ifndef __APPLE__
+    .type MatmulFloatNeon64OptRemain, %function
+#endif
+
+// void MatmulFloatNeon64(const float *a, const float *b, float *c, int depth
+//                        int row, int col, size_t stride)
+// x0: a
+// x1: b
+// x2: c
+// x3: depth
+// x4: row
+// x5: col
+// x6: stride
+// only for winograd
+MatmulFloatNeon64OptRemain:
+    mov x18, #32 // sizeof(float) * 8
+    mul x9, x3, x18 // block stride of lhs/rhs: sizeof(float) * 8 * depth
+    mov x18, #4
+    mul x8, x5, x6
+    mov x11, #8
+    mul x11, x11, x6
+    mul x8, x8, x18
+    mul x11, x11, x18
+
+    cmp x4, #4
+    ble LoopH4
+
+    LoopH8:
+        mov x10, x4 // reload lhs row
+        mov x12, x0 // reload lhs ptr
+        mov x18, x2 // reload dst ptr
+
+        LoopW8:
+            mov x16, x1 // reload rhs ptr
+            mov x13, x3 // reload depth
+            dup v16.4s, wzr
+            dup v17.4s, wzr
+            dup v18.4s, wzr
+            dup v19.4s, wzr
+            dup v20.4s, wzr
+            dup v21.4s, wzr
+            dup v22.4s, wzr
+            dup v23.4s, wzr
+            dup v24.4s, wzr
+            dup v25.4s, wzr
+            dup v26.4s, wzr
+            dup v27.4s, wzr
+            dup v28.4s, wzr
+            dup v29.4s, wzr
+            dup v30.4s, wzr
+            dup v31.4s, wzr
+
+            LoopD8:
+                ld1 {v0.4s, v1.4s, v2.4s}, [x12], #48
+                ld1 {v3.4s, v4.4s}, [x16], #32
+                fmla v16.4s, v3.4s, v0.s[0]
+                fmla v18.4s, v3.4s, v0.s[1]
+                fmla v20.4s, v3.4s, v0.s[2]
+                fmla v22.4s, v3.4s, v0.s[3]
+                fmla v17.4s, v4.4s, v0.s[0]
+                fmla v19.4s, v4.4s, v0.s[1]
+                fmla v21.4s, v4.4s, v0.s[2]
+                fmla v23.4s, v4.4s, v0.s[3]
+                fmla v24.4s, v3.4s, v1.s[0]
+                fmla v26.4s, v3.4s, v1.s[1]
+                fmla v28.4s, v3.4s, v1.s[2]
+                fmla v30.4s, v3.4s, v1.s[3]
+                fmla v25.4s, v4.4s, v1.s[0]
+                fmla v27.4s, v4.4s, v1.s[1]
+                fmla v29.4s, v4.4s, v1.s[2]
+                fmla v31.4s, v4.4s, v1.s[3]
+
+                subs w13, w13, #1
+                bgt LoopD8
+
+                st1 {v16.4s, v17.4s}, [x18], x8
+                st1 {v18.4s, v19.4s}, [x18], x8
+                st1 {v20.4s, v21.4s}, [x18], x8
+                st1 {v22.4s, v23.4s}, [x18], x8
+                st1 {v24.4s, v25.4s}, [x18], x8
+                st1 {v26.4s, v27.4s}, [x18], x8
+                st1 {v28.4s, v29.4s}, [x18], x8
+                st1 {v30.4s, v31.4s}, [x18], x8
+
+            subs x10, x10, #8 // lhs row - 8
+            bgt LoopW8
+
+        subs x5, x5, #8 // rhs col - 8
+        add x1, x1, x9 // rhs ptr + stride
+        add x2, x2, x11
+        bgt LoopH8
+
+        ret
+
+    LoopH4:
+        mov x10, x4 // reload lhs row
+        mov x12, x0 // reload lhs ptr
+        mov x18, x2 // reload dst ptr
+
+        LoopW4:
+            mov x16, x1 // reload rhs ptr
+            mov x13, x3 // reload depth
+            dup v16.4s, wzr
+            dup v17.4s, wzr
+            dup v18.4s, wzr
+            dup v19.4s, wzr
+            dup v20.4s, wzr
+            dup v21.4s, wzr
+            dup v22.4s, wzr
+            dup v23.4s, wzr
+
+            LoopD4:
+                ld1 {v0.4s, v1.4s, v2.4s}, [x12], #48
+                ld1 {v3.4s, v4.4s}, [x16], #32
+                fmla v16.4s, v3.4s, v0.s[0]
+                fmla v18.4s, v3.4s, v0.s[1]
+                fmla v20.4s, v3.4s, v0.s[2]
+                fmla v22.4s, v3.4s, v0.s[3]
+                fmla v17.4s, v4.4s, v0.s[0]
+                fmla v19.4s, v4.4s, v0.s[1]
+                fmla v21.4s, v4.4s, v0.s[2]
+                fmla v23.4s, v4.4s, v0.s[3]
+
+                subs x13, x13, #1
+                bgt LoopD4
+
+                st1 {v16.4s, v17.4s}, [x18], x8
+                st1 {v18.4s, v19.4s}, [x18], x8
+                st1 {v20.4s, v21.4s}, [x18], x8
+                st1 {v22.4s, v23.4s}, [x18], x8
+
+            subs x10, x10, #4 // lhs row - 4
+            bgt LoopW4
+
+        subs x5, x5, #8 // rhs col - 8
+        add x1, x1, x9 // rhs ptr + stride
+        add x2, x2, x11
+        bgt LoopH4
+    ret
+#endif

mindspore/lite/nnacl/fp32/conv.c

@@ -303,7 +303,7 @@ void ConvWinogardFp32(float *input_data, float *trans_weight, const float *bias_
       for (int i = 0; i < input_unit_square; ++i) {
         RowMajor2Col12Major(src_ptr + i * C12NUM * ic4 * C4NUM, tmp_col_ptr, C12NUM, ic4 * C4NUM);
         MatMulOpt(tmp_col_ptr, trans_weight + i * ic4 * C4NUM * oc8 * C8NUM, dst_ptr + i * C8NUM, NULL, 0, ic4 * C4NUM,
-                  C12NUM, oc8 * C8NUM, input_unit_square, 2);
+                  cal_num, oc8 * C8NUM, input_unit_square, 2);
       }
 
       // step 4 : output transform
@@ -489,9 +489,8 @@ void Conv3x3Fp32(float *input_data, float *transed_weight, const float *bias_dat
       for (int i = 0; i < input_unit_square; ++i) {
         RowMajor2Col12Major(src_ptr + i * C12NUM * ic4 * C4NUM, tmp_col_ptr, C12NUM, ic4 * C4NUM);
         MatMulOpt(tmp_col_ptr, transed_weight + i * ic4 * C4NUM * oc8 * C8NUM, dst_ptr + i * C8NUM, NULL, 0,
-                  ic4 * C4NUM, C12NUM, oc8 * C8NUM, input_unit_square, 2);
+                  ic4 * C4NUM, real_cal_num, oc8 * C8NUM, input_unit_square, 2);
       }
-
       Conv3x3Fp32OutputTransform(tmp_dst_buffer + task_id * tmp_dst_buffer_offset, nc4hw4_out + nc4hw4_buffer_offset,
                                  bias_data, start_index, real_cal_num, out_w_block, conv_param);
     }

mindspore/lite/nnacl/fp32/matmul.c

@@ -386,7 +386,7 @@ void MatMul12x8(const float *a, const float *b, float *dst, const float *bias, A
         size_t ci = dst_r_offset + c8div * 8 * stride + c8mod;
         float value = 0;
         for (int d = 0; d < deep; ++d) {
-          size_t ai = src_r_offset + d * row;
+          size_t ai = src_r_offset + d * C12NUM;
           size_t bi = c8div * deep * 8 + d * 8 + c8mod;
           value = value + a[ai] * b[bi];
         }
@@ -403,8 +403,12 @@ void MatMul12x8(const float *a, const float *b, float *dst, const float *bias, A
 void MatMulOpt(const float *a, const float *b, float *c, const float *bias, ActType act_type, int deep, int row,
                int col, size_t stride, int out_type) {
 #ifdef ENABLE_ARM64
-  MatmulFloatNeon64Opt(a, b, c, bias, (int)act_type, deep, row, col, stride, (int)(out_type == OutType_Nhwc),
+  if (out_type == 2 && row <= 8) {
-                       (int)(out_type == OutType_TileC8));
+    MatmulFloatNeon64OptRemain(a, b, c, deep, row, col, stride);
+  } else {
+    MatmulFloatNeon64Opt(a, b, c, bias, (int)act_type, deep, row, col, stride, (int)(out_type == OutType_Nhwc),
+                         (int)(out_type == OutType_TileC8));
+  }
 #else
   MatMul12x8(a, b, c, bias, act_type, deep, row, col, stride, out_type);
 #endif

mindspore/lite/nnacl/fp32/matmul.h

@@ -39,6 +39,7 @@ void MatmulFloatNeon64(const float *a, const float *b, float *c, const float *bi
                        int col, size_t stride, bool write_nhwc);
 void MatmulFloatNeon64Opt(const float *a, const float *b, float *c, const float *bias, int act_type, int depth, int row,
                           int col, size_t stride, size_t write_nhwc, size_t write_c4);
+void MatmulFloatNeon64OptRemain(const float *a, const float *b, float *c, int depth, int row, int col, size_t stride);
 #endif
 #ifdef __cplusplus
 }