diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp | 108 |
1 files changed, 54 insertions, 54 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp index 0dabc2f292..6d2641b035 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -53,21 +53,21 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "b.none 7f\n" "1:" // 4-vectors of channels "lsr x25, %x[n_valid_cells], #0x2\n" + "mov z5.s, #0xff800000\n" "mov z4.s, #0xff800000\n" - "mov z3.s, #0xff800000\n" "mov x24, %x[inptrs]\n" + "mov z3.s, #0xff800000\n" "mov z2.s, #0xff800000\n" - "mov z1.s, #0xff800000\n" "cbz x25, 4f\n" "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "ld1w { z0.s }, p4/Z, [x23, x9, LSL #2]\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" - "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z1.s }, p4/Z, [x23, x9, LSL #2]\n" + "ld1w { z0.s }, p4/Z, [x22, x9, LSL #2]\n" "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" - "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" - "ld1w { z18.s }, p3/Z, [x23, x28, LSL #2]\n" + "ld1w { z31.s }, p4/Z, [x20, x9, LSL #2]\n" + "ld1w { z30.s }, p3/Z, [x23, x28, LSL #2]\n" "ld1w { z29.s }, p3/Z, [x22, x28, LSL #2]\n" "ld1w { z22.s }, p3/Z, [x21, x28, LSL #2]\n" "ld1w { z28.s }, p3/Z, [x20, x28, LSL #2]\n" @@ -81,34 +81,34 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "ld1w { z24.s }, p1/Z, [x20, x26, LSL #2]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop - "movprfx z19, z0\n fmax z19.s, p0/M, z19.s, z31.s\n" - "fmax z23.s, p0/M, z23.s, z30.s\n" + "movprfx z19, z1\n fmax z19.s, p0/M, z19.s, z0.s\n" + "fmax z23.s, p0/M, z23.s, z31.s\n" "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "fmax z18.s, p0/M, z18.s, z29.s\n" + "movprfx z18, z30\n fmax z18.s, p0/M, z18.s, z29.s\n" "fmax z22.s, p0/M, z22.s, z28.s\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" "fmax z17.s, p0/M, z17.s, z27.s\n" "fmax z21.s, p0/M, z21.s, z26.s\n" - "ld1w { z0.s }, p4/Z, [x23, x9, LSL #2]\n" "fmax z16.s, p0/M, z16.s, z25.s\n" "fmax z20.s, p0/M, z20.s, z24.s\n" - "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z1.s }, p4/Z, [x23, x9, LSL #2]\n" "fmax z19.s, p0/M, z19.s, z23.s\n" + "ld1w { z0.s }, p4/Z, [x22, x9, LSL #2]\n" "fmax z18.s, p0/M, z18.s, z22.s\n" "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" "fmax z17.s, p0/M, z17.s, z21.s\n" + "ld1w { z31.s }, p4/Z, [x20, x9, LSL #2]\n" "fmax z16.s, p0/M, z16.s, z20.s\n" - "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" - "fmax z4.s, p0/M, z4.s, z19.s\n" - "fmax z3.s, p0/M, z3.s, z18.s\n" - "ld1w { z18.s }, p3/Z, [x23, x28, LSL #2]\n" - "fmax z2.s, p0/M, z2.s, z17.s\n" - "fmax z1.s, p0/M, z1.s, z16.s\n" + "ld1w { z30.s }, p3/Z, [x23, x28, LSL #2]\n" + "fmax z5.s, p0/M, z5.s, z19.s\n" "ld1w { z29.s }, p3/Z, [x22, x28, LSL #2]\n" + "fmax z4.s, p0/M, z4.s, z18.s\n" "ld1w { z22.s }, p3/Z, [x21, x28, LSL #2]\n" + "fmax z3.s, p0/M, z3.s, z17.s\n" "ld1w { z28.s }, p3/Z, [x20, x28, LSL #2]\n" + "fmax z2.s, p0/M, z2.s, z16.s\n" "ld1w { z17.s }, p2/Z, [x23, x27, LSL #2]\n" "ld1w { z27.s }, p2/Z, [x22, x27, LSL #2]\n" "ld1w { z21.s }, p2/Z, [x21, x27, LSL #2]\n" @@ -119,9 +119,9 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "ld1w { z24.s }, p1/Z, [x20, x26, LSL #2]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail - "movprfx z19, z0\n fmax z19.s, p0/M, z19.s, z31.s\n" - "fmax z23.s, p0/M, z23.s, z30.s\n" - "fmax z18.s, p0/M, z18.s, z29.s\n" + "movprfx z19, z1\n fmax z19.s, p0/M, z19.s, z0.s\n" + "fmax z23.s, p0/M, z23.s, z31.s\n" + "movprfx z18, z30\n fmax z18.s, p0/M, z18.s, z29.s\n" "fmax z22.s, p0/M, z22.s, z28.s\n" "fmax z17.s, p0/M, z17.s, z27.s\n" "fmax z21.s, p0/M, z21.s, z26.s\n" @@ -131,33 +131,33 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "fmax z18.s, p0/M, z18.s, z22.s\n" "fmax z17.s, p0/M, z17.s, z21.s\n" "fmax z16.s, p0/M, z16.s, z20.s\n" - "fmax z4.s, p0/M, z4.s, z19.s\n" - "fmax z3.s, p0/M, z3.s, z18.s\n" - "fmax z2.s, p0/M, z2.s, z17.s\n" - "fmax z1.s, p0/M, z1.s, z16.s\n" + "fmax z5.s, p0/M, z5.s, z19.s\n" + "fmax z4.s, p0/M, z4.s, z18.s\n" + "fmax z3.s, p0/M, z3.s, z17.s\n" + "fmax z2.s, p0/M, z2.s, z16.s\n" "4:" // 4-vectors of channels: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop "ldr x20, [x24], #0x8\n" - "ld1w { z16.s }, p4/Z, [x20, x9, LSL #2]\n" "subs x21, x21, #0x1\n" - "fmax z4.s, p0/M, z4.s, z16.s\n" - "ld1w { z16.s }, p3/Z, [x20, x28, LSL #2]\n" - "fmax z3.s, p0/M, z3.s, z16.s\n" - "ld1w { z16.s }, p2/Z, [x20, x27, LSL #2]\n" - "fmax z2.s, p0/M, z2.s, z16.s\n" + "ld1w { z19.s }, p4/Z, [x20, x9, LSL #2]\n" + "ld1w { z18.s }, p3/Z, [x20, x28, LSL #2]\n" + "ld1w { z17.s }, p2/Z, [x20, x27, LSL #2]\n" "ld1w { z16.s }, p1/Z, [x20, x26, LSL #2]\n" - "fmax z1.s, p0/M, z1.s, z16.s\n" + "fmax z5.s, p0/M, z5.s, z19.s\n" + "fmax z4.s, p0/M, z4.s, z18.s\n" + "fmax z3.s, p0/M, z3.s, z17.s\n" + "fmax z2.s, p0/M, z2.s, z16.s\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "st1w { z4.s }, p4, [%x[outptr], x9, LSL #2]\n" + "st1w { z5.s }, p4, [%x[outptr], x9, LSL #2]\n" "incw x9, ALL, MUL #4\n" - "st1w { z3.s }, p3, [%x[outptr], x28, LSL #2]\n" + "st1w { z4.s }, p3, [%x[outptr], x28, LSL #2]\n" "incw x28, ALL, MUL #4\n" - "st1w { z2.s }, p2, [%x[outptr], x27, LSL #2]\n" + "st1w { z3.s }, p2, [%x[outptr], x27, LSL #2]\n" "incw x27, ALL, MUL #4\n" - "st1w { z1.s }, p1, [%x[outptr], x26, LSL #2]\n" + "st1w { z2.s }, p1, [%x[outptr], x26, LSL #2]\n" "incw x26, ALL, MUL #4\n" "whilelt p1.s, x26, %x[n_channels]\n" "b.any 1b\n" @@ -166,48 +166,48 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "b.none 14f\n" "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" - "mov z4.s, #0xff800000\n" + "mov z5.s, #0xff800000\n" "mov x24, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x20, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "ld1w { z0.s }, p4/Z, [x20, x9, LSL #2]\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" - "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z1.s }, p4/Z, [x23, x9, LSL #2]\n" + "ld1w { z0.s }, p4/Z, [x22, x9, LSL #2]\n" "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" - "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" + "ld1w { z31.s }, p4/Z, [x20, x9, LSL #2]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop - "movprfx z16, z0\n fmax z16.s, p0/M, z16.s, z31.s\n" - "movprfx z17, z23\n fmax z17.s, p0/M, z17.s, z30.s\n" + "movprfx z16, z1\n fmax z16.s, p0/M, z16.s, z0.s\n" + "movprfx z17, z23\n fmax z17.s, p0/M, z17.s, z31.s\n" "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "fmax z16.s, p0/M, z16.s, z17.s\n" "ldp x21, x20, [x24, #0x10]\n" - "fmax z4.s, p0/M, z4.s, z16.s\n" "add x24, x24, #0x20\n" - "ld1w { z0.s }, p4/Z, [x23, x9, LSL #2]\n" - "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z1.s }, p4/Z, [x23, x9, LSL #2]\n" + "fmax z16.s, p0/M, z16.s, z17.s\n" + "ld1w { z0.s }, p4/Z, [x22, x9, LSL #2]\n" "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" - "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" + "ld1w { z31.s }, p4/Z, [x20, x9, LSL #2]\n" + "fmax z5.s, p0/M, z5.s, z16.s\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail - "movprfx z16, z0\n fmax z16.s, p0/M, z16.s, z31.s\n" - "movprfx z17, z23\n fmax z17.s, p0/M, z17.s, z30.s\n" + "movprfx z16, z1\n fmax z16.s, p0/M, z16.s, z0.s\n" + "movprfx z17, z23\n fmax z17.s, p0/M, z17.s, z31.s\n" "fmax z16.s, p0/M, z16.s, z17.s\n" - "fmax z4.s, p0/M, z4.s, z16.s\n" + "fmax z5.s, p0/M, z5.s, z16.s\n" "11:" // Single vector of channels: Loop: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop "ldr x20, [x24], #0x8\n" - "ld1w { z16.s }, p4/Z, [x20, x9, LSL #2]\n" "subs x21, x21, #0x1\n" - "fmax z4.s, p0/M, z4.s, z16.s\n" + "ld1w { z16.s }, p4/Z, [x20, x9, LSL #2]\n" + "fmax z5.s, p0/M, z5.s, z16.s\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "st1w { z4.s }, p4, [%x[outptr], x9, LSL #2]\n" + "st1w { z5.s }, p4, [%x[outptr], x9, LSL #2]\n" "incw x9\n" "whilelt p4.s, x9, %x[n_channels]\n" "b.any 8b\n" |