diff options
author | Michael Tyler <michael.tyler@arm.com> | 2023-04-12 17:43:17 +0100 |
---|---|---|
committer | michael.tyler <michael.tyler@arm.com> | 2023-06-05 15:57:58 +0000 |
commit | 74921eee924625426429044decefe3673561b174 (patch) | |
tree | 654da1a95e3d42d6af8ad1ff27bb40d77b1fd8c5 /src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst | |
parent | df5d9878008be9b60586df97ebfff197abb5195e (diff) | |
download | ComputeLibrary-74921eee924625426429044decefe3673561b174.tar.gz |
Update CPU kernel implementations and guard directives
Resolves COMPMID-6023
Change-Id: I868975d14c4f98af6716726feda22405a6a4c891
Signed-off-by: Michael Tyler <michael.tyler@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp | 146 |
1 files changed, 73 insertions, 73 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp index f888038a2a..f3f4950a1f 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp @@ -109,7 +109,7 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( "lsr x23, %x[n_valid_cells], #0x1\n" "mov z15.s, #0x0\n" "mov z14.s, #0x0\n" - "mov x20, %x[inptrs]\n" + "mov x22, %x[inptrs]\n" "mov z13.s, #0x0\n" "mov z12.s, #0x0\n" "mov z11.s, #0x0\n" @@ -125,42 +125,42 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( "mov z1.s, #0x0\n" "mov z0.s, #0x0\n" "cbz x23, 4f\n" - "ldp x22, x21, [x20, #0x0]\n" + "ldp x21, x20, [x22, #0x0]\n" "subs x23, x23, #0x1\n" - "add x20, x20, #0x10\n" - "ld1b { z31.b }, p4/Z, [x22, x27]\n" - "ld1b { z30.b }, p4/Z, [x21, x27]\n" - "ld1b { z29.b }, p3/Z, [x22, x26]\n" - "ld1b { z28.b }, p3/Z, [x21, x26]\n" - "ld1b { z27.b }, p2/Z, [x22, x25]\n" - "ld1b { z26.b }, p2/Z, [x21, x25]\n" - "ld1b { z25.b }, p1/Z, [x22, x24]\n" - "ld1b { z24.b }, p1/Z, [x21, x24]\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" + "ld1b { z29.b }, p3/Z, [x21, x26]\n" + "ld1b { z28.b }, p3/Z, [x20, x26]\n" + "ld1b { z27.b }, p2/Z, [x21, x25]\n" + "ld1b { z26.b }, p2/Z, [x20, x25]\n" + "ld1b { z25.b }, p1/Z, [x21, x24]\n" + "ld1b { z24.b }, p1/Z, [x20, x24]\n" "beq 3f\n" "2:" // 4-vectors of channels: 2 inputs loop ".inst 0x455e0bf7 // uaddlb z23.h, z31.b, z30.b\n" ".inst 0x455e0ff6 // uaddlt z22.h, z31.b, z30.b\n" - "ldp x22, x21, [x20, #0x0]\n" + "ldp x21, x20, [x22, #0x0]\n" "subs x23, x23, #0x1\n" ".inst 0x455c0bb5 // uaddlb z21.h, z29.b, z28.b\n" ".inst 0x455c0fb4 // uaddlt z20.h, z29.b, z28.b\n" - "add x20, x20, #0x10\n" - "ld1b { z31.b }, p4/Z, [x22, x27]\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" ".inst 0x455a0b73 // uaddlb z19.h, z27.b, z26.b\n" ".inst 0x455a0f72 // uaddlt z18.h, z27.b, z26.b\n" - "ld1b { z30.b }, p4/Z, [x21, x27]\n" - "ld1b { z29.b }, p3/Z, [x22, x26]\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" + "ld1b { z29.b }, p3/Z, [x21, x26]\n" ".inst 0x45580b31 // uaddlb z17.h, z25.b, z24.b\n" ".inst 0x45580f30 // uaddlt z16.h, z25.b, z24.b\n" - "ld1b { z28.b }, p3/Z, [x21, x26]\n" - "ld1b { z27.b }, p2/Z, [x22, x25]\n" + "ld1b { z28.b }, p3/Z, [x20, x26]\n" + "ld1b { z27.b }, p2/Z, [x21, x25]\n" ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" - "ld1b { z26.b }, p2/Z, [x21, x25]\n" - "ld1b { z25.b }, p1/Z, [x22, x24]\n" + "ld1b { z26.b }, p2/Z, [x20, x25]\n" + "ld1b { z25.b }, p1/Z, [x21, x24]\n" ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" - "ld1b { z24.b }, p1/Z, [x21, x24]\n" + "ld1b { z24.b }, p1/Z, [x20, x24]\n" ".inst 0x4595496b // uaddwb z11.s, z11.s, z21.h\n" ".inst 0x45954d4a // uaddwt z10.s, z10.s, z21.h\n" ".inst 0x45944929 // uaddwb z9.s, z9.s, z20.h\n" @@ -203,20 +203,20 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( "ands x21, %x[n_valid_cells], #0x1\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x22, [x20], #0x8\n" - "ld1b { z31.b }, p4/Z, [x22, x27]\n" - ".inst 0x4508abf7 // ushllb z23.h, z31.b, #0x0\n" - ".inst 0x4508aff6 // ushllt z22.h, z31.b, #0x0\n" - "ld1b { z29.b }, p3/Z, [x22, x26]\n" - "ld1b { z27.b }, p2/Z, [x22, x25]\n" - ".inst 0x4508abb5 // ushllb z21.h, z29.b, #0x0\n" - ".inst 0x4508afb4 // ushllt z20.h, z29.b, #0x0\n" - "ld1b { z25.b }, p1/Z, [x22, x24]\n" - ".inst 0x4508ab73 // ushllb z19.h, z27.b, #0x0\n" - ".inst 0x4508af72 // ushllt z18.h, z27.b, #0x0\n" + "ldr x20, [x22], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x27]\n" + ".inst 0x4508aa17 // ushllb z23.h, z16.b, #0x0\n" + ".inst 0x4508ae16 // ushllt z22.h, z16.b, #0x0\n" + "ld1b { z16.b }, p3/Z, [x20, x26]\n" + "ld1b { z17.b }, p2/Z, [x20, x25]\n" + ".inst 0x4508aa15 // ushllb z21.h, z16.b, #0x0\n" + ".inst 0x4508ae14 // ushllt z20.h, z16.b, #0x0\n" + "ld1b { z16.b }, p1/Z, [x20, x24]\n" + ".inst 0x4508aa33 // ushllb z19.h, z17.b, #0x0\n" + ".inst 0x4508ae32 // ushllt z18.h, z17.b, #0x0\n" "subs x21, x21, #0x1\n" - ".inst 0x4508ab31 // ushllb z17.h, z25.b, #0x0\n" - ".inst 0x4508af30 // ushllt z16.h, z25.b, #0x0\n" + ".inst 0x4508aa11 // ushllb z17.h, z16.b, #0x0\n" + ".inst 0x4508ae10 // ushllt z16.h, z16.b, #0x0\n" ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" @@ -332,49 +332,49 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( "lsr x23, %x[n_valid_cells], #0x1\n" "mov z15.s, #0x0\n" "mov z14.s, #0x0\n" - "mov x20, %x[inptrs]\n" + "mov x22, %x[inptrs]\n" "mov z13.s, #0x0\n" "mov z12.s, #0x0\n" "cbz x23, 11f\n" - "ldp x22, x21, [x20, #0x0]\n" + "ldp x21, x20, [x22, #0x0]\n" "subs x23, x23, #0x1\n" - "add x20, x20, #0x10\n" - "ld1b { z31.b }, p4/Z, [x22, x27]\n" - "ld1b { z30.b }, p4/Z, [x21, x27]\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 2 inputs loop - ".inst 0x455e0bf7 // uaddlb z23.h, z31.b, z30.b\n" - ".inst 0x455e0ff6 // uaddlt z22.h, z31.b, z30.b\n" - "ldp x22, x21, [x20, #0x0]\n" + ".inst 0x455e0bf1 // uaddlb z17.h, z31.b, z30.b\n" + ".inst 0x455e0ff0 // uaddlt z16.h, z31.b, z30.b\n" + "ldp x21, x20, [x22, #0x0]\n" "subs x23, x23, #0x1\n" - ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" - ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" - "add x20, x20, #0x10\n" - "ld1b { z31.b }, p4/Z, [x22, x27]\n" - ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" - ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" - "ld1b { z30.b }, p4/Z, [x21, x27]\n" + ".inst 0x459149ef // uaddwb z15.s, z15.s, z17.h\n" + ".inst 0x45914dce // uaddwt z14.s, z14.s, z17.h\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" + ".inst 0x459049ad // uaddwb z13.s, z13.s, z16.h\n" + ".inst 0x45904d8c // uaddwt z12.s, z12.s, z16.h\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 2 inputs tail - ".inst 0x455e0bf7 // uaddlb z23.h, z31.b, z30.b\n" - ".inst 0x455e0ff6 // uaddlt z22.h, z31.b, z30.b\n" - ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" - ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" - ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" - ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" + ".inst 0x455e0bf1 // uaddlb z17.h, z31.b, z30.b\n" + ".inst 0x455e0ff0 // uaddlt z16.h, z31.b, z30.b\n" + ".inst 0x459149ef // uaddwb z15.s, z15.s, z17.h\n" + ".inst 0x45914dce // uaddwt z14.s, z14.s, z17.h\n" + ".inst 0x459049ad // uaddwb z13.s, z13.s, z16.h\n" + ".inst 0x45904d8c // uaddwt z12.s, z12.s, z16.h\n" "11:" // Single vector of channels: Loop: After loop "ands x21, %x[n_valid_cells], #0x1\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x22, [x20], #0x8\n" - "ld1b { z31.b }, p4/Z, [x22, x27]\n" - ".inst 0x4508abf7 // ushllb z23.h, z31.b, #0x0\n" - ".inst 0x4508aff6 // ushllt z22.h, z31.b, #0x0\n" + "ldr x20, [x22], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x27]\n" + ".inst 0x4508aa11 // ushllb z17.h, z16.b, #0x0\n" + ".inst 0x4508ae10 // ushllt z16.h, z16.b, #0x0\n" "subs x21, x21, #0x1\n" - ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" - ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" - ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" - ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" + ".inst 0x459149ef // uaddwb z15.s, z15.s, z17.h\n" + ".inst 0x45914dce // uaddwt z14.s, z14.s, z17.h\n" + ".inst 0x459049ad // uaddwb z13.s, z13.s, z16.h\n" + ".inst 0x45904d8c // uaddwt z12.s, z12.s, z16.h\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End "ld1rw { z17.s }, p0/Z, [%x[rescale_ptr]]\n" @@ -387,17 +387,17 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( ".inst 0x4482820e // srshl z14.s, p0/M, z14.s, z16.s\n" ".inst 0x4482820d // srshl z13.s, p0/M, z13.s, z16.s\n" ".inst 0x4482820c // srshl z12.s, p0/M, z12.s, z16.s\n" - "mov z16.s, #0x0\n" - "mov z18.s, #0xff\n" - "smax z15.s, p0/M, z15.s, z16.s\n" - "smax z14.s, p0/M, z14.s, z16.s\n" - "smax z13.s, p0/M, z13.s, z16.s\n" - "smax z12.s, p0/M, z12.s, z16.s\n" - "smin z15.s, p0/M, z15.s, z18.s\n" - "smin z14.s, p0/M, z14.s, z18.s\n" + "mov z17.s, #0x0\n" + "mov z16.s, #0xff\n" + "smax z15.s, p0/M, z15.s, z17.s\n" + "smax z14.s, p0/M, z14.s, z17.s\n" + "smax z13.s, p0/M, z13.s, z17.s\n" + "smax z12.s, p0/M, z12.s, z17.s\n" + "smin z15.s, p0/M, z15.s, z16.s\n" + "smin z14.s, p0/M, z14.s, z16.s\n" "trn1 z17.h, z15.h, z14.h\n" - "smin z13.s, p0/M, z13.s, z18.s\n" - "smin z12.s, p0/M, z12.s, z18.s\n" + "smin z13.s, p0/M, z13.s, z16.s\n" + "smin z12.s, p0/M, z12.s, z16.s\n" "trn1 z16.h, z13.h, z12.h\n" "trn1 z16.b, z17.b, z16.b\n" "st1b { z16.b }, p4, [%x[outptr], x27]\n" |