diff options
author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-02-02 14:59:09 +0000 |
---|---|---|
committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-02-04 17:43:55 +0000 |
commit | cf87f509fc23d02c56569f794a3fb59e1b8be277 (patch) | |
tree | 0fe55158f2065dc6a314e82935558b9748165285 /src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp | |
parent | 89de118ccbebd5a943634137d0c160d4867da49c (diff) | |
download | ComputeLibrary-cf87f509fc23d02c56569f794a3fb59e1b8be277.tar.gz |
Tweak scheduling use of SQDMULH in quantized AVG pooling
Resolves COMPMID-4195
Change-Id: Ie5116c1ddddccafba40432fd4b5245bb27890a88
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4997
Reviewed-by: TeresaARM <teresa.charlinreyes@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp | 32 |
1 files changed, 16 insertions, 16 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp index 7990a3d9fc..164847480b 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp @@ -84,33 +84,33 @@ void sve_u8_nhwc_max_generic_depthfirst_impl( "subs x24, x24, #0x1\n" "movprfx z23, z1\n umax z23.b, p4/M, z23.b, z0.b\n" "ldp x21, x20, [x19, #0x10]\n" - "movprfx z18, z31\n umax z18.b, p4/M, z18.b, z30.b\n" "add x19, x19, #0x20\n" - "umax z22.b, p4/M, z22.b, z29.b\n" + "movprfx z18, z31\n umax z18.b, p4/M, z18.b, z30.b\n" "ld1b { z3.b }, p3/Z, [x23, x28]\n" + "umax z22.b, p4/M, z22.b, z29.b\n" "movprfx z17, z28\n umax z17.b, p4/M, z17.b, z27.b\n" - "umax z21.b, p4/M, z21.b, z26.b\n" "ld1b { z2.b }, p3/Z, [x22, x28]\n" - "umax z16.b, p4/M, z16.b, z25.b\n" + "umax z21.b, p4/M, z21.b, z26.b\n" "ld1b { z1.b }, p3/Z, [x21, x28]\n" - "umax z20.b, p4/M, z20.b, z24.b\n" + "umax z16.b, p4/M, z16.b, z25.b\n" "ld1b { z0.b }, p3/Z, [x20, x28]\n" - "umax z19.b, p4/M, z19.b, z23.b\n" + "umax z20.b, p4/M, z20.b, z24.b\n" "ld1b { z31.b }, p2/Z, [x23, x27]\n" - "umax z18.b, p4/M, z18.b, z22.b\n" + "umax z19.b, p4/M, z19.b, z23.b\n" "ld1b { z30.b }, p2/Z, [x22, x27]\n" - "umax z17.b, p4/M, z17.b, z21.b\n" + "umax z18.b, p4/M, z18.b, z22.b\n" "ld1b { z22.b }, p2/Z, [x21, x27]\n" - "umax z16.b, p4/M, z16.b, z20.b\n" + "umax z17.b, p4/M, z17.b, z21.b\n" "ld1b { z29.b }, p2/Z, [x20, x27]\n" - "umax z7.b, p4/M, z7.b, z19.b\n" + "umax z16.b, p4/M, z16.b, z20.b\n" "ld1b { z28.b }, p1/Z, [x23, x26]\n" - "umax z6.b, p4/M, z6.b, z18.b\n" + "umax z7.b, p4/M, z7.b, z19.b\n" "ld1b { z27.b }, p1/Z, [x22, x26]\n" - "umax z5.b, p4/M, z5.b, z17.b\n" + "umax z6.b, p4/M, z6.b, z18.b\n" "ld1b { z21.b }, p1/Z, [x21, x26]\n" - "umax z4.b, p4/M, z4.b, z16.b\n" + "umax z5.b, p4/M, z5.b, z17.b\n" "ld1b { z26.b }, p1/Z, [x20, x26]\n" + "umax z4.b, p4/M, z4.b, z16.b\n" "ld1b { z16.b }, p0/Z, [x23, x25]\n" "ld1b { z25.b }, p0/Z, [x22, x25]\n" "ld1b { z20.b }, p0/Z, [x21, x25]\n" @@ -168,9 +168,9 @@ void sve_u8_nhwc_max_generic_depthfirst_impl( "lsr x24, %x[n_valid_cells], #0x2\n" "cbz x24, 11f\n" "ldp x23, x22, [x19, #0x0]\n" - "subs x24, x24, #0x1\n" "ldp x21, x20, [x19, #0x10]\n" "add x19, x19, #0x20\n" + "subs x24, x24, #0x1\n" "ld1b { z3.b }, p3/Z, [x23, x28]\n" "ld1b { z2.b }, p3/Z, [x22, x28]\n" "ld1b { z1.b }, p3/Z, [x21, x28]\n" @@ -182,11 +182,11 @@ void sve_u8_nhwc_max_generic_depthfirst_impl( "subs x24, x24, #0x1\n" "movprfx z23, z1\n umax z23.b, p4/M, z23.b, z0.b\n" "ldp x21, x20, [x19, #0x10]\n" + "add x19, x19, #0x20\n" "umax z19.b, p4/M, z19.b, z23.b\n" "ld1b { z3.b }, p3/Z, [x23, x28]\n" - "umax z7.b, p4/M, z7.b, z19.b\n" "ld1b { z2.b }, p3/Z, [x22, x28]\n" - "add x19, x19, #0x20\n" + "umax z7.b, p4/M, z7.b, z19.b\n" "ld1b { z1.b }, p3/Z, [x21, x28]\n" "ld1b { z0.b }, p3/Z, [x20, x28]\n" "bgt 9b\n" |