diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp | 144 |
1 files changed, 72 insertions, 72 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp index e661e2698a..539b4946b4 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_4VL.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -34,67 +34,67 @@ void sve_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt __asm__ __volatile__( "cmp %x[height], #0x4\n" - "ptrue p4.b\n" + "ptrue p2.b\n" "blt 4f\n" "1:" // Main row loop: Head - "mov x26, %x[in]\n" - "add x25, x26, %x[in_stride]\n" - "add x24, x25, %x[in_stride]\n" - "add x23, x24, %x[in_stride]\n" - "add %x[in], x23, %x[in_stride]\n" - "mov x22, %x[out]\n" + "mov x27, %x[in]\n" + "mov x26, %x[out]\n" "sub %x[height], %x[height], #0x4\n" - "mov x21, %x[width]\n" + "mov x25, %x[width]\n" + "add x24, x27, %x[in_stride]\n" + "add x23, x24, %x[in_stride]\n" + "add x22, x23, %x[in_stride]\n" + "add %x[in], x22, %x[in_stride]\n" "2:" // Main row loop: Column loop - "mov x20, x21\n" - "whilelt p3.h, XZR, x20\n" - "ld1h { z31.h }, p3/Z, [x26]\n" - "ld1h { z30.h }, p3/Z, [x25]\n" - "dech x20\n" - "whilelt p2.h, XZR, x20\n" - "ld1h { z29.h }, p2/Z, [x26, #1, MUL VL]\n" - "ld1h { z28.h }, p2/Z, [x25, #1, MUL VL]\n" - "dech x20\n" - "whilelt p1.h, XZR, x20\n" - "ld1h { z27.h }, p1/Z, [x26, #2, MUL VL]\n" - "ld1h { z26.h }, p1/Z, [x25, #2, MUL VL]\n" - "dech x20\n" - "whilelt p0.h, XZR, x20\n" - "ld1h { z25.h }, p0/Z, [x26, #3, MUL VL]\n" - "ld1h { z24.h }, p0/Z, [x25, #3, MUL VL]\n" - "mov x20, x22\n" - "dech x21, ALL, MUL #4\n" - "ld1h { z23.h }, p3/Z, [x24]\n" - "ld1h { z22.h }, p2/Z, [x24, #1, MUL VL]\n" - "ld1h { z21.h }, p1/Z, [x24, #2, MUL VL]\n" - "ld1h { z20.h }, p0/Z, [x24, #3, MUL VL]\n" - "cmp x21, #0x0\n" - "addvl x26, x26, #4\n" - "ld1h { z19.h }, p3/Z, [x23]\n" - "ld1h { z18.h }, p2/Z, [x23, #1, MUL VL]\n" - "addvl x25, x25, #4\n" + "mov x21, x25\n" + "mov x20, x26\n" + "dech x25, ALL, MUL #4\n" + "add x26, x26, %x[out_stride]\n" + "whilelt p1.h, XZR, x21\n" + "dech x21\n" + "whilelt p0.h, XZR, x21\n" + "dech x21\n" + "ld1h { z31.h }, p1/Z, [x27]\n" + "ld1h { z30.h }, p1/Z, [x24]\n" + "ld1h { z29.h }, p1/Z, [x23]\n" + "ld1h { z28.h }, p1/Z, [x22]\n" + "whilelt p1.h, XZR, x21\n" + "dech x21\n" + "ld1h { z27.h }, p0/Z, [x27, #1, MUL VL]\n" + "ld1h { z26.h }, p0/Z, [x24, #1, MUL VL]\n" + "ld1h { z25.h }, p0/Z, [x23, #1, MUL VL]\n" + "ld1h { z24.h }, p0/Z, [x22, #1, MUL VL]\n" + "whilelt p0.h, XZR, x21\n" + "cmp x25, #0x0\n" + "ld1h { z23.h }, p1/Z, [x27, #2, MUL VL]\n" + "ld1h { z22.h }, p1/Z, [x24, #2, MUL VL]\n" + "ld1h { z21.h }, p1/Z, [x23, #2, MUL VL]\n" + "ld1h { z20.h }, p1/Z, [x22, #2, MUL VL]\n" + "ld1h { z19.h }, p0/Z, [x27, #3, MUL VL]\n" + "ld1h { z18.h }, p0/Z, [x24, #3, MUL VL]\n" + "addvl x27, x27, #4\n" "addvl x24, x24, #4\n" - "ld1h { z17.h }, p1/Z, [x23, #2, MUL VL]\n" - "ld1h { z16.h }, p0/Z, [x23, #3, MUL VL]\n" - "st1h { z31.h }, p4, [x20]\n" + "ld1h { z17.h }, p0/Z, [x23, #3, MUL VL]\n" + "ld1h { z16.h }, p0/Z, [x22, #3, MUL VL]\n" + "st1h { z31.h }, p2, [x20]\n" "addvl x23, x23, #4\n" - "st1h { z29.h }, p4, [x20, #1, MUL VL]\n" - "add x22, x22, %x[out_stride]\n" - "st1h { z27.h }, p4, [x20, #2, MUL VL]\n" - "st1h { z25.h }, p4, [x20, #3, MUL VL]\n" - "st1h { z30.h }, p4, [x20, #4, MUL VL]\n" - "st1h { z28.h }, p4, [x20, #5, MUL VL]\n" - "st1h { z26.h }, p4, [x20, #6, MUL VL]\n" - "st1h { z24.h }, p4, [x20, #7, MUL VL]\n" + "st1h { z27.h }, p2, [x20, #1, MUL VL]\n" + "addvl x22, x22, #4\n" + "st1h { z23.h }, p2, [x20, #2, MUL VL]\n" + "st1h { z19.h }, p2, [x20, #3, MUL VL]\n" + "st1h { z30.h }, p2, [x20, #4, MUL VL]\n" + "st1h { z26.h }, p2, [x20, #5, MUL VL]\n" + "st1h { z22.h }, p2, [x20, #6, MUL VL]\n" + "st1h { z18.h }, p2, [x20, #7, MUL VL]\n" "addvl x20, x20, #16\n" - "st1h { z23.h }, p4, [x20, #-8, MUL VL]\n" - "st1h { z22.h }, p4, [x20, #-7, MUL VL]\n" - "st1h { z21.h }, p4, [x20, #-6, MUL VL]\n" - "st1h { z20.h }, p4, [x20, #-5, MUL VL]\n" - "st1h { z19.h }, p4, [x20, #-4, MUL VL]\n" - "st1h { z18.h }, p4, [x20, #-3, MUL VL]\n" - "st1h { z17.h }, p4, [x20, #-2, MUL VL]\n" - "st1h { z16.h }, p4, [x20, #-1, MUL VL]\n" + "st1h { z29.h }, p2, [x20, #-8, MUL VL]\n" + "st1h { z25.h }, p2, [x20, #-7, MUL VL]\n" + "st1h { z21.h }, p2, [x20, #-6, MUL VL]\n" + "st1h { z17.h }, p2, [x20, #-5, MUL VL]\n" + "st1h { z28.h }, p2, [x20, #-4, MUL VL]\n" + "st1h { z24.h }, p2, [x20, #-3, MUL VL]\n" + "st1h { z20.h }, p2, [x20, #-2, MUL VL]\n" + "st1h { z16.h }, p2, [x20, #-1, MUL VL]\n" "bgt 2b\n" "3:" // Main row loop: Column loop skip "cmp %x[height], #0x4\n" @@ -103,32 +103,32 @@ void sve_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt "cbz %x[height], 8f\n" "4:" // Main loop skip "5:" // Tail row loop: Head - "mov x26, %x[in]\n" - "add %x[in], x26, %x[in_stride]\n" - "mov x22, %x[out]\n" + "mov x27, %x[in]\n" + "mov x26, %x[out]\n" "sub %x[height], %x[height], #0x1\n" "mov x21, %x[width]\n" + "add %x[in], x27, %x[in_stride]\n" "6:" // Tail row loop: Column loop "mov x20, x21\n" - "whilelt p0.h, XZR, x20\n" - "ld1h { z19.h }, p0/Z, [x26]\n" + "dech x21, ALL, MUL #4\n" + "whilelt p1.h, XZR, x20\n" "dech x20\n" "whilelt p0.h, XZR, x20\n" - "ld1h { z18.h }, p0/Z, [x26, #1, MUL VL]\n" "dech x20\n" - "whilelt p0.h, XZR, x20\n" - "ld1h { z17.h }, p0/Z, [x26, #2, MUL VL]\n" + "ld1h { z19.h }, p1/Z, [x27]\n" + "whilelt p1.h, XZR, x20\n" "dech x20\n" - "dech x21, ALL, MUL #4\n" + "ld1h { z18.h }, p0/Z, [x27, #1, MUL VL]\n" "whilelt p0.h, XZR, x20\n" "cmp x21, #0x0\n" - "ld1h { z16.h }, p0/Z, [x26, #3, MUL VL]\n" - "st1h { z19.h }, p4, [x22]\n" - "addvl x26, x26, #4\n" - "st1h { z18.h }, p4, [x22, #1, MUL VL]\n" - "st1h { z17.h }, p4, [x22, #2, MUL VL]\n" - "st1h { z16.h }, p4, [x22, #3, MUL VL]\n" - "add x22, x22, %x[out_stride]\n" + "ld1h { z17.h }, p1/Z, [x27, #2, MUL VL]\n" + "ld1h { z16.h }, p0/Z, [x27, #3, MUL VL]\n" + "addvl x27, x27, #4\n" + "st1h { z19.h }, p2, [x26]\n" + "st1h { z18.h }, p2, [x26, #1, MUL VL]\n" + "st1h { z17.h }, p2, [x26, #2, MUL VL]\n" + "st1h { z16.h }, p2, [x26, #3, MUL VL]\n" + "add x26, x26, %x[out_stride]\n" "bgt 6b\n" "7:" // Tail row loop: Column loop skip "cmp %x[height], #0x1\n" @@ -137,7 +137,7 @@ void sve_transpose_interleave_4VL(uint16_t *out, const uint16_t *in, size_t widt "8:" // Done : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out) : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } |