diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-01-14 13:43:40 +0000 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-01-18 12:39:30 +0000 |
commit | 33e03074c36d85de87e9032a2583b04ce8ddcd6b (patch) | |
tree | 4442ec5b5022fa5681d689e6ccf3a6423efa8a93 /src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL | |
parent | 8d5337ef18901f1b54d0c062ae7486bc5a4c6610 (diff) | |
download | ComputeLibrary-33e03074c36d85de87e9032a2583b04ce8ddcd6b.tar.gz |
Cycle estimate-based kernel selection for dot product quantized s8/u8 kernels
Resolves: COMPMID-3990
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: If840c79209940535450f4ea1cbf6b0ec646a168e
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4866
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp | 55 |
1 files changed, 14 insertions, 41 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp index b19842b122..bb42dc0e04 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -328,7 +328,6 @@ void sve_hybrid_fp16_mla_6x4VL ( "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" "cmp x12, x19\n" "bne 7b\n" - "prfm pstl1keep, [x13, #0x0]\n" "tbz %x[flags], #1, 13f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rh { z1.h }, p5/Z, [x19]\n" @@ -349,9 +348,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "st1h { z11.h }, p1, [x13, #3, MUL VL]\n" "addvl x13, x13, #4\n" "14:" // Height 1: Writeback done - "mov x19, #0x0\n" - "inch x19, ALL, MUL #4\n" - "subs x16, x16, x19\n" + "dech x16, ALL, MUL #4\n" + "cmp x16, XZR\n" "bgt 3b\n" "b 86f\n" "15:" // Height 2 @@ -670,8 +668,6 @@ void sve_hybrid_fp16_mla_6x4VL ( "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" "cmp x12, x19\n" "bne 21b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" "tbz %x[flags], #1, 27f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rh { z1.h }, p5/Z, [x19]\n" @@ -705,9 +701,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "st1h { z15.h }, p1, [x9, #3, MUL VL]\n" "addvl x9, x9, #4\n" "28:" // Height 2: Writeback done - "mov x19, #0x0\n" - "inch x19, ALL, MUL #4\n" - "subs x16, x16, x19\n" + "dech x16, ALL, MUL #4\n" + "cmp x16, XZR\n" "bgt 17b\n" "b 86f\n" "29:" // Height 3 @@ -1114,9 +1109,6 @@ void sve_hybrid_fp16_mla_6x4VL ( "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" "cmp x12, x19\n" "bne 35b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" "tbz %x[flags], #1, 41f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rh { z1.h }, p5/Z, [x19]\n" @@ -1163,9 +1155,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "st1h { z19.h }, p1, [x27, #3, MUL VL]\n" "addvl x27, x27, #4\n" "42:" // Height 3: Writeback done - "mov x19, #0x0\n" - "inch x19, ALL, MUL #4\n" - "subs x16, x16, x19\n" + "dech x16, ALL, MUL #4\n" + "cmp x16, XZR\n" "bgt 31b\n" "b 86f\n" "43:" // Height 4 @@ -1660,10 +1651,6 @@ void sve_hybrid_fp16_mla_6x4VL ( "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" "cmp x12, x19\n" "bne 49b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" "tbz %x[flags], #1, 55f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rh { z1.h }, p5/Z, [x19]\n" @@ -1723,9 +1710,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "st1h { z23.h }, p1, [x25, #3, MUL VL]\n" "addvl x25, x25, #4\n" "56:" // Height 4: Writeback done - "mov x19, #0x0\n" - "inch x19, ALL, MUL #4\n" - "subs x16, x16, x19\n" + "dech x16, ALL, MUL #4\n" + "cmp x16, XZR\n" "bgt 45b\n" "b 86f\n" "57:" // Height 5 @@ -2308,11 +2294,6 @@ void sve_hybrid_fp16_mla_6x4VL ( "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" "cmp x12, x19\n" "bne 63b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" - "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 69f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rh { z1.h }, p5/Z, [x19]\n" @@ -2385,9 +2366,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "st1h { z27.h }, p1, [x23, #3, MUL VL]\n" "addvl x23, x23, #4\n" "70:" // Height 5: Writeback done - "mov x19, #0x0\n" - "inch x19, ALL, MUL #4\n" - "subs x16, x16, x19\n" + "dech x16, ALL, MUL #4\n" + "cmp x16, XZR\n" "bgt 59b\n" "b 86f\n" "71:" // Height 6 @@ -3060,12 +3040,6 @@ void sve_hybrid_fp16_mla_6x4VL ( "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" "cmp x12, x19\n" "bne 77b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" - "prfm pstl1keep, [x23, #0x0]\n" - "prfm pstl1keep, [x21, #0x0]\n" "tbz %x[flags], #1, 83f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1rh { z1.h }, p5/Z, [x19]\n" @@ -3151,9 +3125,8 @@ void sve_hybrid_fp16_mla_6x4VL ( "st1h { z31.h }, p1, [x21, #3, MUL VL]\n" "addvl x21, x21, #4\n" "84:" // Height 6: Writeback done - "mov x19, #0x0\n" - "inch x19, ALL, MUL #4\n" - "subs x16, x16, x19\n" + "dech x16, ALL, MUL #4\n" + "cmp x16, XZR\n" "bgt 73b\n" "subs %x[M], %x[M], #0x6\n" "beq 86f\n" @@ -3168,7 +3141,7 @@ void sve_hybrid_fp16_mla_6x4VL ( "b 1b\n" "86:" // Exit - : [M] "+r" (M), [input_ptr] "+r" (input_ptr), [output_ptr] "+r" (output_ptr) + : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); |