diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_3VLx8/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_3VLx8/generic.cpp | 31 |
1 files changed, 16 insertions, 15 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_3VLx8/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_3VLx8/generic.cpp index 7640fcaa20..cdc70705c5 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_3VLx8/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_s8s32_dot_3VLx8/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 Arm Limited. + * Copyright (c) 2019-2020 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -51,20 +51,20 @@ void sve_interleaved_s8s32_dot_3VLx8(const int8_t *Apanel, const int8_t *Bpanel, "mov z9.s, #0\n" "mov z10.s, #0\n" "mov z11.s, #0\n" - "mov z12.s, #0\n" "ld1rqb z0.b, p0/z, [%[a_ptr]]\n" - "mov z13.s, #0\n" + "mov z12.s, #0\n" "ld1b z4.b, p0/z, [%[b_ptr]]\n" - "mov z14.s, #0\n" + "mov z13.s, #0\n" "ld1rqb z1.b, p0/z, [%[a_ptr], #0x10]\n" - "mov z15.s, #0\n" + "mov z14.s, #0\n" "ld1b z5.b, p0/z, [%[b_ptr], #1, MUL VL]\n" - "mov z16.s, #0\n" + "mov z15.s, #0\n" "ld1rqb z2.b, p0/z, [%[a_ptr], #0x20]\n" - "mov z17.s, #0\n" + "mov z16.s, #0\n" "add %[a_ptr], %[a_ptr], #0x40\n" - "mov z18.s, #0\n" + "mov z17.s, #0\n" "addvl %[b_ptr], %[b_ptr], #3\n" + "mov z18.s, #0\n" "mov z19.s, #0\n" "mov z20.s, #0\n" "mov z21.s, #0\n" @@ -208,8 +208,8 @@ void sve_interleaved_s8s32_dot_3VLx8(const int8_t *Apanel, const int8_t *Bpanel, "sdot z9.s, z4.b, z0.b[1]\n" "sdot z10.s, z4.b, z0.b[2]\n" "sdot z11.s, z4.b, z0.b[3]\n" - "sdot z20.s, z4.b, z1.b[0]\n" "st1w z8.s, p0, [%[c_ptr]]\n" + "sdot z20.s, z4.b, z1.b[0]\n" "sdot z21.s, z4.b, z1.b[1]\n" "sdot z22.s, z4.b, z1.b[2]\n" "sdot z23.s, z4.b, z1.b[3]\n" @@ -217,8 +217,8 @@ void sve_interleaved_s8s32_dot_3VLx8(const int8_t *Apanel, const int8_t *Bpanel, "sdot z13.s, z5.b, z0.b[1]\n" "sdot z14.s, z5.b, z0.b[2]\n" "sdot z15.s, z5.b, z0.b[3]\n" - "sdot z24.s, z5.b, z1.b[0]\n" "st1w z12.s, p0, [%[c_ptr], #1, MUL VL]\n" + "sdot z24.s, z5.b, z1.b[0]\n" "sdot z25.s, z5.b, z1.b[1]\n" "sdot z26.s, z5.b, z1.b[2]\n" "sdot z27.s, z5.b, z1.b[3]\n" @@ -226,10 +226,11 @@ void sve_interleaved_s8s32_dot_3VLx8(const int8_t *Apanel, const int8_t *Bpanel, "sdot z17.s, z6.b, z0.b[1]\n" "sdot z18.s, z6.b, z0.b[2]\n" "sdot z19.s, z6.b, z0.b[3]\n" - "sdot z28.s, z6.b, z1.b[0]\n" "st1w z16.s, p0, [%[c_ptr], #2, MUL VL]\n" + "sdot z28.s, z6.b, z1.b[0]\n" "sdot z29.s, z6.b, z1.b[1]\n" "sdot z30.s, z6.b, z1.b[2]\n" + "st1w z9.s, p0, [%[c_ptr], #3, MUL VL]\n" "sdot z31.s, z6.b, z1.b[3]\n" "b 4f\n" "3:\n" @@ -267,8 +268,8 @@ void sve_interleaved_s8s32_dot_3VLx8(const int8_t *Apanel, const int8_t *Bpanel, "sdot z9.s, z4.b, z2.b[1]\n" "sdot z10.s, z4.b, z2.b[2]\n" "sdot z11.s, z4.b, z2.b[3]\n" - "sdot z20.s, z4.b, z3.b[0]\n" "st1w z8.s, p0, [%[c_ptr]]\n" + "sdot z20.s, z4.b, z3.b[0]\n" "sdot z21.s, z4.b, z3.b[1]\n" "sdot z22.s, z4.b, z3.b[2]\n" "sdot z23.s, z4.b, z3.b[3]\n" @@ -276,8 +277,8 @@ void sve_interleaved_s8s32_dot_3VLx8(const int8_t *Apanel, const int8_t *Bpanel, "sdot z13.s, z5.b, z2.b[1]\n" "sdot z14.s, z5.b, z2.b[2]\n" "sdot z15.s, z5.b, z2.b[3]\n" - "sdot z24.s, z5.b, z3.b[0]\n" "st1w z12.s, p0, [%[c_ptr], #1, MUL VL]\n" + "sdot z24.s, z5.b, z3.b[0]\n" "sdot z25.s, z5.b, z3.b[1]\n" "sdot z26.s, z5.b, z3.b[2]\n" "sdot z27.s, z5.b, z3.b[3]\n" @@ -285,13 +286,13 @@ void sve_interleaved_s8s32_dot_3VLx8(const int8_t *Apanel, const int8_t *Bpanel, "sdot z17.s, z6.b, z2.b[1]\n" "sdot z18.s, z6.b, z2.b[2]\n" "sdot z19.s, z6.b, z2.b[3]\n" - "sdot z28.s, z6.b, z3.b[0]\n" "st1w z16.s, p0, [%[c_ptr], #2, MUL VL]\n" + "sdot z28.s, z6.b, z3.b[0]\n" "sdot z29.s, z6.b, z3.b[1]\n" "sdot z30.s, z6.b, z3.b[2]\n" + "st1w z9.s, p0, [%[c_ptr], #3, MUL VL]\n" "sdot z31.s, z6.b, z3.b[3]\n" "4:\n" - "st1w z9.s, p0, [%[c_ptr], #3, MUL VL]\n" "st1w z13.s, p0, [%[c_ptr], #4, MUL VL]\n" "st1w z17.s, p0, [%[c_ptr], #5, MUL VL]\n" "st1w z10.s, p0, [%[c_ptr], #6, MUL VL]\n" |