diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp | 98 |
1 files changed, 49 insertions, 49 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp index 60f17b7bc2..dee5b4a230 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp16_nhwc_avg_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -49,25 +49,25 @@ void sme_fp16_nhwc_avg_generic_depthfirst_impl( "cnth x26, ALL, MUL #3\n" "ptrue p0.b\n" "whilelt p3.h, x9, %x[n_channels]\n" - "ld1rh { z6.h }, p0/Z, [%x[rescale_ptr]]\n" "whilelt p2.h, x28, %x[n_channels]\n" + "ld1rh { z5.h }, p0/Z, [%x[rescale_ptr]]\n" "whilelt p1.h, x27, %x[n_channels]\n" "whilelt p0.h, x26, %x[n_channels]\n" "b.none 7f\n" "1:" // 4-vectors of channels "lsr x25, %x[n_valid_cells], #0x2\n" - "mov z5.b, #0x0\n" "mov z4.b, #0x0\n" - "mov x24, %x[inptrs]\n" "mov z3.b, #0x0\n" + "mov x24, %x[inptrs]\n" "mov z2.b, #0x0\n" + "mov z1.b, #0x0\n" "cbz x25, 4f\n" "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "ld1h { z1.h }, p3/Z, [x23, x9, LSL #1]\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" - "ld1h { z0.h }, p3/Z, [x22, x9, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x23, x9, LSL #1]\n" + "ld1h { z23.h }, p3/Z, [x22, x9, LSL #1]\n" "ld1h { z31.h }, p3/Z, [x21, x9, LSL #1]\n" "ld1h { z30.h }, p3/Z, [x20, x9, LSL #1]\n" "ld1h { z29.h }, p2/Z, [x23, x28, LSL #1]\n" @@ -84,7 +84,7 @@ void sme_fp16_nhwc_avg_generic_depthfirst_impl( "ld1h { z16.h }, p0/Z, [x20, x26, LSL #1]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop - "fadd z23.h, z1.h, z0.h\n" + "fadd z23.h, z0.h, z23.h\n" "fadd z19.h, z31.h, z30.h\n" "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" @@ -94,24 +94,24 @@ void sme_fp16_nhwc_avg_generic_depthfirst_impl( "add x24, x24, #0x20\n" "fadd z21.h, z27.h, z21.h\n" "fadd z17.h, z26.h, z17.h\n" - "ld1h { z1.h }, p3/Z, [x23, x9, LSL #1]\n" "fadd z20.h, z25.h, z20.h\n" "fadd z16.h, z24.h, z16.h\n" - "ld1h { z0.h }, p3/Z, [x22, x9, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x23, x9, LSL #1]\n" "fadd z19.h, z23.h, z19.h\n" + "ld1h { z23.h }, p3/Z, [x22, x9, LSL #1]\n" "fadd z18.h, z22.h, z18.h\n" "ld1h { z31.h }, p3/Z, [x21, x9, LSL #1]\n" "fadd z17.h, z21.h, z17.h\n" - "fadd z16.h, z20.h, z16.h\n" "ld1h { z30.h }, p3/Z, [x20, x9, LSL #1]\n" - "fadd z5.h, z5.h, z19.h\n" - "fadd z4.h, z4.h, z18.h\n" + "fadd z16.h, z20.h, z16.h\n" "ld1h { z29.h }, p2/Z, [x23, x28, LSL #1]\n" - "fadd z3.h, z3.h, z17.h\n" - "fadd z2.h, z2.h, z16.h\n" + "fadd z4.h, z4.h, z19.h\n" "ld1h { z22.h }, p2/Z, [x22, x28, LSL #1]\n" + "fadd z3.h, z3.h, z18.h\n" "ld1h { z28.h }, p2/Z, [x21, x28, LSL #1]\n" + "fadd z2.h, z2.h, z17.h\n" "ld1h { z18.h }, p2/Z, [x20, x28, LSL #1]\n" + "fadd z1.h, z1.h, z16.h\n" "ld1h { z27.h }, p1/Z, [x23, x27, LSL #1]\n" "ld1h { z21.h }, p1/Z, [x22, x27, LSL #1]\n" "ld1h { z26.h }, p1/Z, [x21, x27, LSL #1]\n" @@ -122,7 +122,7 @@ void sme_fp16_nhwc_avg_generic_depthfirst_impl( "ld1h { z16.h }, p0/Z, [x20, x26, LSL #1]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail - "fadd z23.h, z1.h, z0.h\n" + "fadd z23.h, z0.h, z23.h\n" "fadd z19.h, z31.h, z30.h\n" "fadd z22.h, z29.h, z22.h\n" "fadd z18.h, z28.h, z18.h\n" @@ -134,37 +134,37 @@ void sme_fp16_nhwc_avg_generic_depthfirst_impl( "fadd z18.h, z22.h, z18.h\n" "fadd z17.h, z21.h, z17.h\n" "fadd z16.h, z20.h, z16.h\n" - "fadd z5.h, z5.h, z19.h\n" - "fadd z4.h, z4.h, z18.h\n" - "fadd z3.h, z3.h, z17.h\n" - "fadd z2.h, z2.h, z16.h\n" + "fadd z4.h, z4.h, z19.h\n" + "fadd z3.h, z3.h, z18.h\n" + "fadd z2.h, z2.h, z17.h\n" + "fadd z1.h, z1.h, z16.h\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" - "ld1h { z16.h }, p3/Z, [x20, x9, LSL #1]\n" "subs x21, x21, #0x1\n" - "fadd z5.h, z5.h, z16.h\n" - "ld1h { z16.h }, p2/Z, [x20, x28, LSL #1]\n" - "fadd z4.h, z4.h, z16.h\n" - "ld1h { z16.h }, p1/Z, [x20, x27, LSL #1]\n" - "fadd z3.h, z3.h, z16.h\n" + "ld1h { z19.h }, p3/Z, [x20, x9, LSL #1]\n" + "ld1h { z18.h }, p2/Z, [x20, x28, LSL #1]\n" + "ld1h { z17.h }, p1/Z, [x20, x27, LSL #1]\n" "ld1h { z16.h }, p0/Z, [x20, x26, LSL #1]\n" - "fadd z2.h, z2.h, z16.h\n" + "fadd z4.h, z4.h, z19.h\n" + "fadd z3.h, z3.h, z18.h\n" + "fadd z2.h, z2.h, z17.h\n" + "fadd z1.h, z1.h, z16.h\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "fmul z5.h, z5.h, z6.h\n" - "fmul z4.h, z4.h, z6.h\n" - "st1h { z5.h }, p3, [%x[outptr], x9, LSL #1]\n" + "fmul z4.h, z4.h, z5.h\n" + "fmul z3.h, z3.h, z5.h\n" + "fmul z2.h, z2.h, z5.h\n" + "fmul z1.h, z1.h, z5.h\n" + "st1h { z4.h }, p3, [%x[outptr], x9, LSL #1]\n" "inch x9, ALL, MUL #4\n" - "fmul z3.h, z3.h, z6.h\n" - "fmul z2.h, z2.h, z6.h\n" - "st1h { z4.h }, p2, [%x[outptr], x28, LSL #1]\n" + "st1h { z3.h }, p2, [%x[outptr], x28, LSL #1]\n" "inch x28, ALL, MUL #4\n" - "st1h { z3.h }, p1, [%x[outptr], x27, LSL #1]\n" + "st1h { z2.h }, p1, [%x[outptr], x27, LSL #1]\n" "inch x27, ALL, MUL #4\n" - "st1h { z2.h }, p0, [%x[outptr], x26, LSL #1]\n" + "st1h { z1.h }, p0, [%x[outptr], x26, LSL #1]\n" "inch x26, ALL, MUL #4\n" "whilelt p0.h, x26, %x[n_channels]\n" "b.any 1b\n" @@ -173,49 +173,49 @@ void sme_fp16_nhwc_avg_generic_depthfirst_impl( "b.none 14f\n" "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" - "mov z5.b, #0x0\n" + "mov z4.b, #0x0\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" - "ld1h { z1.h }, p3/Z, [x20, x9, LSL #1]\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" - "ld1h { z0.h }, p3/Z, [x22, x9, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x23, x9, LSL #1]\n" + "ld1h { z23.h }, p3/Z, [x22, x9, LSL #1]\n" "ld1h { z31.h }, p3/Z, [x21, x9, LSL #1]\n" "ld1h { z30.h }, p3/Z, [x20, x9, LSL #1]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop - "fadd z17.h, z1.h, z0.h\n" + "fadd z17.h, z0.h, z23.h\n" "fadd z16.h, z31.h, z30.h\n" "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "fadd z16.h, z17.h, z16.h\n" "ldp x21, x20, [x24, #0x10]\n" - "fadd z5.h, z5.h, z16.h\n" "add x24, x24, #0x20\n" - "ld1h { z1.h }, p3/Z, [x23, x9, LSL #1]\n" - "ld1h { z0.h }, p3/Z, [x22, x9, LSL #1]\n" + "ld1h { z0.h }, p3/Z, [x23, x9, LSL #1]\n" + "fadd z16.h, z17.h, z16.h\n" + "ld1h { z23.h }, p3/Z, [x22, x9, LSL #1]\n" "ld1h { z31.h }, p3/Z, [x21, x9, LSL #1]\n" "ld1h { z30.h }, p3/Z, [x20, x9, LSL #1]\n" + "fadd z4.h, z4.h, z16.h\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail - "fadd z17.h, z1.h, z0.h\n" + "fadd z17.h, z0.h, z23.h\n" "fadd z16.h, z31.h, z30.h\n" "fadd z16.h, z17.h, z16.h\n" - "fadd z5.h, z5.h, z16.h\n" + "fadd z4.h, z4.h, z16.h\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" - "ld1h { z16.h }, p3/Z, [x20, x9, LSL #1]\n" "subs x21, x21, #0x1\n" - "fadd z5.h, z5.h, z16.h\n" + "ld1h { z16.h }, p3/Z, [x20, x9, LSL #1]\n" + "fadd z4.h, z4.h, z16.h\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "fmul z5.h, z5.h, z6.h\n" - "st1h { z5.h }, p3, [%x[outptr], x9, LSL #1]\n" + "fmul z4.h, z4.h, z5.h\n" + "st1h { z4.h }, p3, [%x[outptr], x9, LSL #1]\n" "inch x9\n" "whilelt p3.h, x9, %x[n_channels]\n" "b.any 8b\n" |