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_s8_nhwc_max_2x2_s1_output2x2_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_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp | 32 |
1 files changed, 16 insertions, 16 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp index 06c777b8ed..bdf3f53292 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp @@ -65,13 +65,13 @@ void sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst_impl( __asm__ __volatile__( "ldr x14, [%x[args], %[offsetof_n_channels]]\n" "ptrue p2.b\n" - "ldr x19, [%x[args], %[offsetof_outptrs]]\n" + "ldr x20, [%x[args], %[offsetof_outptrs]]\n" "mov x13, #0x0\n" + "ldr x19, [%x[args], %[offsetof_inptrs]]\n" "mov x12, #0x0\n" - "ldp x11, x10, [x19, #0x0]\n" + "ldp x11, x10, [x20, #0x0]\n" "whilelt p1.b, x13, x14\n" - "ldp x9, x28, [x19, #0x10]\n" - "ldr x19, [%x[args], %[offsetof_inptrs]]\n" + "ldp x9, x28, [x20, #0x10]\n" "ldp x27, x26, [x19, #0x0]\n" "ldp x25, x24, [x19, #0x10]\n" "ldp x23, x22, [x19, #0x20]\n" @@ -97,20 +97,20 @@ void sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst_impl( "ld1b { z30.b }, p1/Z, [x23, x13]\n" "movprfx z18, z28\n smax z18.b, p2/M, z18.b, z27.b\n" "ld1b { z29.b }, p1/Z, [x20, x13]\n" - "movprfx z20, z26\n smax z20.b, p2/M, z20.b, z25.b\n" + "movprfx z17, z26\n smax z17.b, p2/M, z17.b, z25.b\n" "ld1b { z27.b }, p1/Z, [x27, x13]\n" - "movprfx z17, z24\n smax z17.b, p2/M, z17.b, z28.b\n" + "movprfx z16, z24\n smax z16.b, p2/M, z16.b, z28.b\n" "ld1b { z28.b }, p1/Z, [x24, x13]\n" - "movprfx z16, z26\n smax z16.b, p2/M, z16.b, z23.b\n" + "movprfx z20, z26\n smax z20.b, p2/M, z20.b, z23.b\n" "ld1b { z26.b }, p1/Z, [x22, x13]\n" "movprfx z19, z22\n smax z19.b, p2/M, z19.b, z18.b\n" "ld1b { z25.b }, p1/Z, [x25, x13]\n" - "movprfx z18, z22\n smax z18.b, p2/M, z18.b, z20.b\n" + "movprfx z18, z22\n smax z18.b, p2/M, z18.b, z17.b\n" "ld1b { z24.b }, p1/Z, [x21, x13]\n" - "smax z17.b, p2/M, z17.b, z21.b\n" + "movprfx z17, z21\n smax z17.b, p2/M, z17.b, z16.b\n" "ld1b { z23.b }, p1/Z, [x19, x13]\n" "incw x13\n" - "smax z16.b, p2/M, z16.b, z21.b\n" + "movprfx z16, z21\n smax z16.b, p2/M, z16.b, z20.b\n" "st1b { z19.b }, p0, [x11, x12]\n" "whilelt p1.b, x13, x14\n" "st1b { z18.b }, p0, [x10, x12]\n" @@ -123,15 +123,15 @@ void sve_s8_nhwc_max_2x2_s1_output2x2_depthfirst_impl( "whilelt p0.b, x12, x14\n" "movprfx z21, z30\n smax z21.b, p2/M, z21.b, z29.b\n" "movprfx z18, z28\n smax z18.b, p2/M, z18.b, z27.b\n" - "movprfx z20, z26\n smax z20.b, p2/M, z20.b, z25.b\n" - "movprfx z17, z24\n smax z17.b, p2/M, z17.b, z28.b\n" - "movprfx z16, z26\n smax z16.b, p2/M, z16.b, z23.b\n" + "movprfx z17, z26\n smax z17.b, p2/M, z17.b, z25.b\n" + "movprfx z16, z24\n smax z16.b, p2/M, z16.b, z28.b\n" + "movprfx z20, z26\n smax z20.b, p2/M, z20.b, z23.b\n" "movprfx z19, z22\n smax z19.b, p2/M, z19.b, z18.b\n" "st1b { z19.b }, p0, [x11, x12]\n" - "movprfx z18, z22\n smax z18.b, p2/M, z18.b, z20.b\n" - "smax z17.b, p2/M, z17.b, z21.b\n" + "movprfx z18, z22\n smax z18.b, p2/M, z18.b, z17.b\n" + "movprfx z17, z21\n smax z17.b, p2/M, z17.b, z16.b\n" "st1b { z18.b }, p0, [x10, x12]\n" - "smax z16.b, p2/M, z16.b, z21.b\n" + "movprfx z16, z21\n smax z16.b, p2/M, z16.b, z20.b\n" "st1b { z17.b }, p0, [x9, x12]\n" "st1b { z16.b }, p0, [x28, x12]\n" : |