diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp | 144 |
1 files changed, 72 insertions, 72 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp index 36bfccf52f..1e3f2f300b 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp16_mla_8x24/generic.cpp @@ -51,27 +51,27 @@ void a64_ffinterleaved_fp16_mla_8x24( __asm__ __volatile__( "1:" // Height loop - "ldr x25, [%x[args_ptr], %[offsetof_Bpanel]]\n" - "ldr x24, [%x[args_ptr], %[offsetof_N]]\n" - "str x25, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "mov x23, %x[Apanel]\n" + "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n" + "ldr x25, [%x[args_ptr], %[offsetof_N]]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "mov x24, %x[Apanel]\n" "2:" // Width loop - "ldr x25, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" + "ldr x23, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" - "add x22, x25, x20, LSL #1\n" + "add x22, x23, x20, LSL #1\n" "add x21, x22, x20, LSL #1\n" "add x20, x21, x20, LSL #1\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "cmp x24, #0x10\n" - "mov %x[Apanel], x23\n" + "cmp x25, #0x10\n" + "mov %x[Apanel], x24\n" "bgt 3f\n" - "cmp x24, #0x8\n" - "mov x21, x25\n" + "cmp x25, #0x8\n" + "mov x21, x23\n" "bgt 3f\n" - "mov x22, x25\n" + "mov x22, x23\n" "3:" // B setup done "ldr q0, [%x[Apanel], #0x0]\n" - "ldr q2, [x25, #0x0]\n" + "ldr q2, [x23, #0x0]\n" "movi v8.16b, #0x0\n" "ldr q3, [x22, #0x0]\n" "ldr q4, [x21, #0x0]\n" @@ -102,11 +102,11 @@ void a64_ffinterleaved_fp16_mla_8x24( "movi v31.16b, #0x0\n" "blt 5f\n" "4:" // main loop head - "ldr q1, [%x[Apanel], #0x10]\n" - "ldr q5, [x25, #0x10]\n" + "ldr q7, [%x[Apanel], #0x10]\n" + "ldr q6, [x23, #0x10]\n" "fmla v8.8h, v2.8h, v0.h[0]\n" - "ldr q6, [x22, #0x10]\n" - "ldr q7, [x21, #0x10]\n" + "ldr q5, [x22, #0x10]\n" + "ldr q1, [x21, #0x10]\n" "fmla v11.8h, v2.8h, v0.h[1]\n" "fmla v14.8h, v2.8h, v0.h[2]\n" "fmla v17.8h, v2.8h, v0.h[3]\n" @@ -119,8 +119,8 @@ void a64_ffinterleaved_fp16_mla_8x24( "add %x[Apanel], %x[Apanel], #0x20\n" "fmla v9.8h, v3.8h, v0.h[0]\n" "fmla v12.8h, v3.8h, v0.h[1]\n" - "add x25, x25, #0x20\n" - "ldr q2, [x25, #0x0]\n" + "add x23, x23, #0x20\n" + "ldr q2, [x23, #0x0]\n" "fmla v15.8h, v3.8h, v0.h[2]\n" "fmla v18.8h, v3.8h, v0.h[3]\n" "fmla v21.8h, v3.8h, v0.h[4]\n" @@ -140,30 +140,30 @@ void a64_ffinterleaved_fp16_mla_8x24( "fmla v31.8h, v4.8h, v0.h[7]\n" "ldr q0, [%x[Apanel], #0x0]\n" "ldr q4, [x21, #0x0]\n" - "fmla v8.8h, v5.8h, v1.h[0]\n" - "fmla v11.8h, v5.8h, v1.h[1]\n" - "fmla v14.8h, v5.8h, v1.h[2]\n" - "fmla v17.8h, v5.8h, v1.h[3]\n" - "fmla v20.8h, v5.8h, v1.h[4]\n" - "fmla v23.8h, v5.8h, v1.h[5]\n" - "fmla v26.8h, v5.8h, v1.h[6]\n" - "fmla v29.8h, v5.8h, v1.h[7]\n" - "fmla v9.8h, v6.8h, v1.h[0]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "fmla v15.8h, v6.8h, v1.h[2]\n" - "fmla v18.8h, v6.8h, v1.h[3]\n" - "fmla v21.8h, v6.8h, v1.h[4]\n" - "fmla v24.8h, v6.8h, v1.h[5]\n" - "fmla v27.8h, v6.8h, v1.h[6]\n" - "fmla v30.8h, v6.8h, v1.h[7]\n" - "fmla v10.8h, v7.8h, v1.h[0]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "fmla v16.8h, v7.8h, v1.h[2]\n" - "fmla v19.8h, v7.8h, v1.h[3]\n" - "fmla v22.8h, v7.8h, v1.h[4]\n" - "fmla v25.8h, v7.8h, v1.h[5]\n" - "fmla v28.8h, v7.8h, v1.h[6]\n" - "fmla v31.8h, v7.8h, v1.h[7]\n" + "fmla v8.8h, v6.8h, v7.h[0]\n" + "fmla v11.8h, v6.8h, v7.h[1]\n" + "fmla v14.8h, v6.8h, v7.h[2]\n" + "fmla v17.8h, v6.8h, v7.h[3]\n" + "fmla v20.8h, v6.8h, v7.h[4]\n" + "fmla v23.8h, v6.8h, v7.h[5]\n" + "fmla v26.8h, v6.8h, v7.h[6]\n" + "fmla v29.8h, v6.8h, v7.h[7]\n" + "fmla v9.8h, v5.8h, v7.h[0]\n" + "fmla v12.8h, v5.8h, v7.h[1]\n" + "fmla v15.8h, v5.8h, v7.h[2]\n" + "fmla v18.8h, v5.8h, v7.h[3]\n" + "fmla v21.8h, v5.8h, v7.h[4]\n" + "fmla v24.8h, v5.8h, v7.h[5]\n" + "fmla v27.8h, v5.8h, v7.h[6]\n" + "fmla v30.8h, v5.8h, v7.h[7]\n" + "fmla v10.8h, v1.8h, v7.h[0]\n" + "fmla v13.8h, v1.8h, v7.h[1]\n" + "fmla v16.8h, v1.8h, v7.h[2]\n" + "fmla v19.8h, v1.8h, v7.h[3]\n" + "fmla v22.8h, v1.8h, v7.h[4]\n" + "fmla v25.8h, v1.8h, v7.h[5]\n" + "fmla v28.8h, v1.8h, v7.h[6]\n" + "fmla v31.8h, v1.8h, v7.h[7]\n" "bge 4b\n" "5:" // main loop skip "fmla v8.8h, v2.8h, v0.h[0]\n" @@ -171,7 +171,7 @@ void a64_ffinterleaved_fp16_mla_8x24( "add %x[Apanel], %x[Apanel], #0x10\n" "fmla v14.8h, v2.8h, v0.h[2]\n" "fmla v17.8h, v2.8h, v0.h[3]\n" - "add x25, x25, #0x10\n" + "add x23, x23, #0x10\n" "fmla v20.8h, v2.8h, v0.h[4]\n" "fmla v23.8h, v2.8h, v0.h[5]\n" "add x22, x22, #0x10\n" @@ -195,37 +195,37 @@ void a64_ffinterleaved_fp16_mla_8x24( "fmla v28.8h, v4.8h, v0.h[6]\n" "fmla v31.8h, v4.8h, v0.h[7]\n" "cbz x20, 6f\n" - "ldr q0, [%x[Apanel], #0x0]\n" - "ldr q5, [x25, #0x0]\n" - "fmla v8.8h, v5.8h, v0.h[0]\n" - "ldr q6, [x22, #0x0]\n" - "ldr q7, [x21, #0x0]\n" - "fmla v11.8h, v5.8h, v0.h[1]\n" - "fmla v14.8h, v5.8h, v0.h[2]\n" - "fmla v17.8h, v5.8h, v0.h[3]\n" + "ldr q3, [%x[Apanel], #0x0]\n" + "ldr q2, [x23, #0x0]\n" + "fmla v8.8h, v2.8h, v3.h[0]\n" + "ldr q1, [x22, #0x0]\n" + "ldr q0, [x21, #0x0]\n" + "fmla v11.8h, v2.8h, v3.h[1]\n" + "fmla v14.8h, v2.8h, v3.h[2]\n" + "fmla v17.8h, v2.8h, v3.h[3]\n" "add %x[Apanel], %x[Apanel], #0x10\n" - "fmla v20.8h, v5.8h, v0.h[4]\n" - "fmla v23.8h, v5.8h, v0.h[5]\n" - "fmla v26.8h, v5.8h, v0.h[6]\n" - "fmla v29.8h, v5.8h, v0.h[7]\n" - "fmla v9.8h, v6.8h, v0.h[0]\n" - "fmla v12.8h, v6.8h, v0.h[1]\n" - "fmla v15.8h, v6.8h, v0.h[2]\n" - "fmla v18.8h, v6.8h, v0.h[3]\n" - "fmla v21.8h, v6.8h, v0.h[4]\n" - "fmla v24.8h, v6.8h, v0.h[5]\n" - "fmla v27.8h, v6.8h, v0.h[6]\n" - "fmla v30.8h, v6.8h, v0.h[7]\n" - "fmla v10.8h, v7.8h, v0.h[0]\n" - "fmla v13.8h, v7.8h, v0.h[1]\n" - "fmla v16.8h, v7.8h, v0.h[2]\n" - "fmla v19.8h, v7.8h, v0.h[3]\n" - "fmla v22.8h, v7.8h, v0.h[4]\n" - "fmla v25.8h, v7.8h, v0.h[5]\n" - "fmla v28.8h, v7.8h, v0.h[6]\n" - "fmla v31.8h, v7.8h, v0.h[7]\n" + "fmla v20.8h, v2.8h, v3.h[4]\n" + "fmla v23.8h, v2.8h, v3.h[5]\n" + "fmla v26.8h, v2.8h, v3.h[6]\n" + "fmla v29.8h, v2.8h, v3.h[7]\n" + "fmla v9.8h, v1.8h, v3.h[0]\n" + "fmla v12.8h, v1.8h, v3.h[1]\n" + "fmla v15.8h, v1.8h, v3.h[2]\n" + "fmla v18.8h, v1.8h, v3.h[3]\n" + "fmla v21.8h, v1.8h, v3.h[4]\n" + "fmla v24.8h, v1.8h, v3.h[5]\n" + "fmla v27.8h, v1.8h, v3.h[6]\n" + "fmla v30.8h, v1.8h, v3.h[7]\n" + "fmla v10.8h, v0.8h, v3.h[0]\n" + "fmla v13.8h, v0.8h, v3.h[1]\n" + "fmla v16.8h, v0.8h, v3.h[2]\n" + "fmla v19.8h, v0.8h, v3.h[3]\n" + "fmla v22.8h, v0.8h, v3.h[4]\n" + "fmla v25.8h, v0.8h, v3.h[5]\n" + "fmla v28.8h, v0.8h, v3.h[6]\n" + "fmla v31.8h, v0.8h, v3.h[7]\n" "6:" // multiply loop done - "subs x24, x24, #0x18\n" + "subs x25, x25, #0x18\n" "str q8, [%x[Cpanel], #0x0]\n" "str q9, [%x[Cpanel], #0x10]\n" "str q10, [%x[Cpanel], #0x20]\n" |