diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp | 222 |
1 files changed, 111 insertions, 111 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp index ec99d64f4a..6de0a380eb 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_fp32_mla_8x12/generic.cpp @@ -51,29 +51,29 @@ void a64_ffinterleaved_fp32_mla_8x12( __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 #2\n" + "add x22, x23, x20, LSL #2\n" "add x21, x22, x20, LSL #2\n" "add x20, x21, x20, LSL #2\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "cmp x24, #0x8\n" - "mov %x[Apanel], x23\n" + "cmp x25, #0x8\n" + "mov %x[Apanel], x24\n" "bgt 3f\n" - "cmp x24, #0x4\n" - "mov x21, x25\n" + "cmp x25, #0x4\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 q1, [%x[Apanel], #0x10]\n" "movi v8.16b, #0x0\n" - "ldr q4, [x25, #0x0]\n" + "ldr q4, [x23, #0x0]\n" "ldr q5, [x22, #0x0]\n" "movi v9.16b, #0x0\n" "ldr q6, [x21, #0x0]\n" @@ -103,10 +103,10 @@ void a64_ffinterleaved_fp32_mla_8x12( "movi v31.16b, #0x0\n" "blt 5f\n" "4:" // main loop head - "ldr q2, [%x[Apanel], #0x20]\n" - "ldr q3, [%x[Apanel], #0x30]\n" + "ldr q3, [%x[Apanel], #0x20]\n" + "ldr q7, [%x[Apanel], #0x30]\n" "fmla v8.4s, v4.4s, v0.s[0]\n" - "ldr q7, [x25, #0x10]\n" + "ldr q2, [x23, #0x10]\n" "fmla v11.4s, v4.4s, v0.s[1]\n" "fmla v14.4s, v4.4s, v0.s[2]\n" "fmla v17.4s, v4.4s, v0.s[3]\n" @@ -136,36 +136,36 @@ void a64_ffinterleaved_fp32_mla_8x12( "fmla v28.4s, v6.4s, v1.s[2]\n" "fmla v31.4s, v6.4s, v1.s[3]\n" "ldr q1, [%x[Apanel], #0x50]\n" - "ldr q6, [x25, #0x20]\n" - "fmla v8.4s, v7.4s, v2.s[0]\n" - "fmla v11.4s, v7.4s, v2.s[1]\n" - "fmla v14.4s, v7.4s, v2.s[2]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "fmla v20.4s, v7.4s, v3.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[1]\n" - "fmla v26.4s, v7.4s, v3.s[2]\n" - "fmla v29.4s, v7.4s, v3.s[3]\n" - "ldr q7, [x22, #0x20]\n" - "fmla v9.4s, v4.4s, v2.s[0]\n" - "fmla v12.4s, v4.4s, v2.s[1]\n" - "fmla v15.4s, v4.4s, v2.s[2]\n" - "fmla v18.4s, v4.4s, v2.s[3]\n" - "fmla v21.4s, v4.4s, v3.s[0]\n" - "fmla v24.4s, v4.4s, v3.s[1]\n" - "fmla v27.4s, v4.4s, v3.s[2]\n" - "fmla v30.4s, v4.4s, v3.s[3]\n" + "ldr q6, [x23, #0x20]\n" + "fmla v8.4s, v2.4s, v3.s[0]\n" + "fmla v11.4s, v2.4s, v3.s[1]\n" + "fmla v14.4s, v2.4s, v3.s[2]\n" + "fmla v17.4s, v2.4s, v3.s[3]\n" + "fmla v20.4s, v2.4s, v7.s[0]\n" + "fmla v23.4s, v2.4s, v7.s[1]\n" + "fmla v26.4s, v2.4s, v7.s[2]\n" + "fmla v29.4s, v2.4s, v7.s[3]\n" + "ldr q2, [x22, #0x20]\n" + "fmla v9.4s, v4.4s, v3.s[0]\n" + "fmla v12.4s, v4.4s, v3.s[1]\n" + "fmla v15.4s, v4.4s, v3.s[2]\n" + "fmla v18.4s, v4.4s, v3.s[3]\n" + "fmla v21.4s, v4.4s, v7.s[0]\n" + "fmla v24.4s, v4.4s, v7.s[1]\n" + "fmla v27.4s, v4.4s, v7.s[2]\n" + "fmla v30.4s, v4.4s, v7.s[3]\n" "ldr q4, [x21, #0x20]\n" - "fmla v10.4s, v5.4s, v2.s[0]\n" - "fmla v13.4s, v5.4s, v2.s[1]\n" - "fmla v16.4s, v5.4s, v2.s[2]\n" - "fmla v19.4s, v5.4s, v2.s[3]\n" - "ldr q2, [%x[Apanel], #0x60]\n" - "fmla v22.4s, v5.4s, v3.s[0]\n" - "fmla v25.4s, v5.4s, v3.s[1]\n" - "fmla v28.4s, v5.4s, v3.s[2]\n" - "fmla v31.4s, v5.4s, v3.s[3]\n" - "ldr q3, [%x[Apanel], #0x70]\n" - "ldr q5, [x25, #0x30]\n" + "fmla v10.4s, v5.4s, v3.s[0]\n" + "fmla v13.4s, v5.4s, v3.s[1]\n" + "fmla v16.4s, v5.4s, v3.s[2]\n" + "fmla v19.4s, v5.4s, v3.s[3]\n" + "ldr q3, [%x[Apanel], #0x60]\n" + "fmla v22.4s, v5.4s, v7.s[0]\n" + "fmla v25.4s, v5.4s, v7.s[1]\n" + "fmla v28.4s, v5.4s, v7.s[2]\n" + "fmla v31.4s, v5.4s, v7.s[3]\n" + "ldr q7, [%x[Apanel], #0x70]\n" + "ldr q5, [x23, #0x30]\n" "fmla v8.4s, v6.4s, v0.s[0]\n" "fmla v11.4s, v6.4s, v0.s[1]\n" "fmla v14.4s, v6.4s, v0.s[2]\n" @@ -173,20 +173,20 @@ void a64_ffinterleaved_fp32_mla_8x12( "add %x[Apanel], %x[Apanel], #0x80\n" "fmla v20.4s, v6.4s, v1.s[0]\n" "fmla v23.4s, v6.4s, v1.s[1]\n" - "add x25, x25, #0x40\n" + "add x23, x23, #0x40\n" "fmla v26.4s, v6.4s, v1.s[2]\n" "fmla v29.4s, v6.4s, v1.s[3]\n" "ldr q6, [x22, #0x30]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v12.4s, v7.4s, v0.s[1]\n" + "fmla v9.4s, v2.4s, v0.s[0]\n" + "fmla v12.4s, v2.4s, v0.s[1]\n" "add x22, x22, #0x40\n" - "fmla v15.4s, v7.4s, v0.s[2]\n" - "fmla v18.4s, v7.4s, v0.s[3]\n" - "fmla v21.4s, v7.4s, v1.s[0]\n" - "fmla v24.4s, v7.4s, v1.s[1]\n" - "fmla v27.4s, v7.4s, v1.s[2]\n" - "fmla v30.4s, v7.4s, v1.s[3]\n" - "ldr q7, [x21, #0x30]\n" + "fmla v15.4s, v2.4s, v0.s[2]\n" + "fmla v18.4s, v2.4s, v0.s[3]\n" + "fmla v21.4s, v2.4s, v1.s[0]\n" + "fmla v24.4s, v2.4s, v1.s[1]\n" + "fmla v27.4s, v2.4s, v1.s[2]\n" + "fmla v30.4s, v2.4s, v1.s[3]\n" + "ldr q2, [x21, #0x30]\n" "fmla v10.4s, v4.4s, v0.s[0]\n" "fmla v13.4s, v4.4s, v0.s[1]\n" "add x21, x21, #0x40\n" @@ -198,33 +198,33 @@ void a64_ffinterleaved_fp32_mla_8x12( "fmla v28.4s, v4.4s, v1.s[2]\n" "fmla v31.4s, v4.4s, v1.s[3]\n" "ldr q1, [%x[Apanel], #0x10]\n" - "ldr q4, [x25, #0x0]\n" - "fmla v8.4s, v5.4s, v2.s[0]\n" - "fmla v11.4s, v5.4s, v2.s[1]\n" - "fmla v14.4s, v5.4s, v2.s[2]\n" - "fmla v17.4s, v5.4s, v2.s[3]\n" - "fmla v20.4s, v5.4s, v3.s[0]\n" - "fmla v23.4s, v5.4s, v3.s[1]\n" - "fmla v26.4s, v5.4s, v3.s[2]\n" - "fmla v29.4s, v5.4s, v3.s[3]\n" + "ldr q4, [x23, #0x0]\n" + "fmla v8.4s, v5.4s, v3.s[0]\n" + "fmla v11.4s, v5.4s, v3.s[1]\n" + "fmla v14.4s, v5.4s, v3.s[2]\n" + "fmla v17.4s, v5.4s, v3.s[3]\n" + "fmla v20.4s, v5.4s, v7.s[0]\n" + "fmla v23.4s, v5.4s, v7.s[1]\n" + "fmla v26.4s, v5.4s, v7.s[2]\n" + "fmla v29.4s, v5.4s, v7.s[3]\n" "ldr q5, [x22, #0x0]\n" - "fmla v9.4s, v6.4s, v2.s[0]\n" - "fmla v12.4s, v6.4s, v2.s[1]\n" - "fmla v15.4s, v6.4s, v2.s[2]\n" - "fmla v18.4s, v6.4s, v2.s[3]\n" - "fmla v21.4s, v6.4s, v3.s[0]\n" - "fmla v24.4s, v6.4s, v3.s[1]\n" - "fmla v27.4s, v6.4s, v3.s[2]\n" - "fmla v30.4s, v6.4s, v3.s[3]\n" + "fmla v9.4s, v6.4s, v3.s[0]\n" + "fmla v12.4s, v6.4s, v3.s[1]\n" + "fmla v15.4s, v6.4s, v3.s[2]\n" + "fmla v18.4s, v6.4s, v3.s[3]\n" + "fmla v21.4s, v6.4s, v7.s[0]\n" + "fmla v24.4s, v6.4s, v7.s[1]\n" + "fmla v27.4s, v6.4s, v7.s[2]\n" + "fmla v30.4s, v6.4s, v7.s[3]\n" "ldr q6, [x21, #0x0]\n" - "fmla v10.4s, v7.4s, v2.s[0]\n" - "fmla v13.4s, v7.4s, v2.s[1]\n" - "fmla v16.4s, v7.4s, v2.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[3]\n" - "fmla v22.4s, v7.4s, v3.s[0]\n" - "fmla v25.4s, v7.4s, v3.s[1]\n" - "fmla v28.4s, v7.4s, v3.s[2]\n" - "fmla v31.4s, v7.4s, v3.s[3]\n" + "fmla v10.4s, v2.4s, v3.s[0]\n" + "fmla v13.4s, v2.4s, v3.s[1]\n" + "fmla v16.4s, v2.4s, v3.s[2]\n" + "fmla v19.4s, v2.4s, v3.s[3]\n" + "fmla v22.4s, v2.4s, v7.s[0]\n" + "fmla v25.4s, v2.4s, v7.s[1]\n" + "fmla v28.4s, v2.4s, v7.s[2]\n" + "fmla v31.4s, v2.4s, v7.s[3]\n" "bge 4b\n" "5:" // main loop skip "fmla v8.4s, v4.4s, v0.s[0]\n" @@ -232,7 +232,7 @@ void a64_ffinterleaved_fp32_mla_8x12( "add %x[Apanel], %x[Apanel], #0x20\n" "fmla v14.4s, v4.4s, v0.s[2]\n" "fmla v17.4s, v4.4s, v0.s[3]\n" - "add x25, x25, #0x10\n" + "add x23, x23, #0x10\n" "fmla v20.4s, v4.4s, v1.s[0]\n" "fmla v23.4s, v4.4s, v1.s[1]\n" "add x22, x22, #0x10\n" @@ -257,43 +257,43 @@ void a64_ffinterleaved_fp32_mla_8x12( "fmla v31.4s, v6.4s, v1.s[3]\n" "cbz x20, 7f\n" "6:" // odd loop - "ldr q0, [%x[Apanel], #0x0]\n" - "ldr q1, [%x[Apanel], #0x10]\n" + "ldr q4, [%x[Apanel], #0x0]\n" + "ldr q3, [%x[Apanel], #0x10]\n" "subs x20, x20, #0x1\n" - "ldr q7, [x25, #0x0]\n" - "ldr q4, [x22, #0x0]\n" - "fmla v8.4s, v7.4s, v0.s[0]\n" - "ldr q5, [x21, #0x0]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v14.4s, v7.4s, v0.s[2]\n" - "fmla v17.4s, v7.4s, v0.s[3]\n" - "fmla v20.4s, v7.4s, v1.s[0]\n" + "ldr q2, [x23, #0x0]\n" + "ldr q1, [x22, #0x0]\n" + "fmla v8.4s, v2.4s, v4.s[0]\n" + "ldr q0, [x21, #0x0]\n" + "fmla v11.4s, v2.4s, v4.s[1]\n" + "fmla v14.4s, v2.4s, v4.s[2]\n" + "fmla v17.4s, v2.4s, v4.s[3]\n" + "fmla v20.4s, v2.4s, v3.s[0]\n" "add %x[Apanel], %x[Apanel], #0x20\n" - "fmla v23.4s, v7.4s, v1.s[1]\n" - "fmla v26.4s, v7.4s, v1.s[2]\n" - "add x25, x25, #0x10\n" - "fmla v29.4s, v7.4s, v1.s[3]\n" - "fmla v9.4s, v4.4s, v0.s[0]\n" + "fmla v23.4s, v2.4s, v3.s[1]\n" + "fmla v26.4s, v2.4s, v3.s[2]\n" + "add x23, x23, #0x10\n" + "fmla v29.4s, v2.4s, v3.s[3]\n" + "fmla v9.4s, v1.4s, v4.s[0]\n" "add x22, x22, #0x10\n" - "fmla v12.4s, v4.4s, v0.s[1]\n" - "fmla v15.4s, v4.4s, v0.s[2]\n" + "fmla v12.4s, v1.4s, v4.s[1]\n" + "fmla v15.4s, v1.4s, v4.s[2]\n" "add x21, x21, #0x10\n" - "fmla v18.4s, v4.4s, v0.s[3]\n" - "fmla v21.4s, v4.4s, v1.s[0]\n" - "fmla v24.4s, v4.4s, v1.s[1]\n" - "fmla v27.4s, v4.4s, v1.s[2]\n" - "fmla v30.4s, v4.4s, v1.s[3]\n" - "fmla v10.4s, v5.4s, v0.s[0]\n" - "fmla v13.4s, v5.4s, v0.s[1]\n" - "fmla v16.4s, v5.4s, v0.s[2]\n" - "fmla v19.4s, v5.4s, v0.s[3]\n" - "fmla v22.4s, v5.4s, v1.s[0]\n" - "fmla v25.4s, v5.4s, v1.s[1]\n" - "fmla v28.4s, v5.4s, v1.s[2]\n" - "fmla v31.4s, v5.4s, v1.s[3]\n" + "fmla v18.4s, v1.4s, v4.s[3]\n" + "fmla v21.4s, v1.4s, v3.s[0]\n" + "fmla v24.4s, v1.4s, v3.s[1]\n" + "fmla v27.4s, v1.4s, v3.s[2]\n" + "fmla v30.4s, v1.4s, v3.s[3]\n" + "fmla v10.4s, v0.4s, v4.s[0]\n" + "fmla v13.4s, v0.4s, v4.s[1]\n" + "fmla v16.4s, v0.4s, v4.s[2]\n" + "fmla v19.4s, v0.4s, v4.s[3]\n" + "fmla v22.4s, v0.4s, v3.s[0]\n" + "fmla v25.4s, v0.4s, v3.s[1]\n" + "fmla v28.4s, v0.4s, v3.s[2]\n" + "fmla v31.4s, v0.4s, v3.s[3]\n" "bne 6b\n" "7:" // multiply loop done - "subs x24, x24, #0xc\n" + "subs x25, x25, #0xc\n" "str q8, [%x[Cpanel], #0x0]\n" "str q9, [%x[Cpanel], #0x10]\n" "str q10, [%x[Cpanel], #0x20]\n" |