diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp | 212 |
1 files changed, 106 insertions, 106 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp index 5ef141492f..19209811d8 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,82 +42,82 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( ) { __asm__ __volatile__( - "mov x28, #0x0\n" - "cntb x27\n" - "cntb x26, ALL, MUL #2\n" - "cntb x25, ALL, MUL #3\n" - "whilelt p4.b, x28, %x[n_channels]\n" - "whilelt p3.b, x27, %x[n_channels]\n" - "whilelt p2.b, x26, %x[n_channels]\n" - "whilelt p1.b, x25, %x[n_channels]\n" + "mov x9, #0x0\n" + "cntb x28\n" + "cntb x27, ALL, MUL #2\n" + "cntb x26, ALL, MUL #3\n" + "whilelt p4.b, x9, %x[n_channels]\n" + "whilelt p3.b, x28, %x[n_channels]\n" + "whilelt p2.b, x27, %x[n_channels]\n" + "whilelt p1.b, x26, %x[n_channels]\n" "ptrue p0.b\n" "b.none 7f\n" "1:" // 4-vectors of channels - "lsr x24, %x[n_valid_cells], #0x2\n" + "lsr x25, %x[n_valid_cells], #0x2\n" "mov z8.b, #0x80\n" "mov z7.b, #0x80\n" - "mov x19, %x[inptrs]\n" + "mov x20, %x[inptrs]\n" "mov z6.b, #0x80\n" "mov z5.b, #0x80\n" - "cbz x24, 4f\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" - "subs x24, x24, #0x1\n" - "add x19, x19, #0x20\n" - "ld1b { z4.b }, p4/Z, [x23, x28]\n" - "ld1b { z3.b }, p4/Z, [x22, x28]\n" - "ld1b { z2.b }, p4/Z, [x21, x28]\n" - "ld1b { z1.b }, p4/Z, [x20, x28]\n" - "ld1b { z0.b }, p3/Z, [x23, x27]\n" - "ld1b { z31.b }, p3/Z, [x22, x27]\n" - "ld1b { z22.b }, p3/Z, [x21, x27]\n" - "ld1b { z30.b }, p3/Z, [x20, x27]\n" - "ld1b { z29.b }, p2/Z, [x23, x26]\n" - "ld1b { z28.b }, p2/Z, [x22, x26]\n" - "ld1b { z21.b }, p2/Z, [x21, x26]\n" - "ld1b { z27.b }, p2/Z, [x20, x26]\n" - "ld1b { z26.b }, p1/Z, [x23, x25]\n" - "ld1b { z25.b }, p1/Z, [x22, x25]\n" - "ld1b { z20.b }, p1/Z, [x21, x25]\n" - "ld1b { z24.b }, p1/Z, [x20, x25]\n" + "cbz x25, 4f\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" + "subs x25, x25, #0x1\n" + "add x20, x20, #0x20\n" + "ld1b { z4.b }, p4/Z, [x24, x9]\n" + "ld1b { z3.b }, p4/Z, [x23, x9]\n" + "ld1b { z2.b }, p4/Z, [x22, x9]\n" + "ld1b { z1.b }, p4/Z, [x21, x9]\n" + "ld1b { z0.b }, p3/Z, [x24, x28]\n" + "ld1b { z31.b }, p3/Z, [x23, x28]\n" + "ld1b { z22.b }, p3/Z, [x22, x28]\n" + "ld1b { z30.b }, p3/Z, [x21, x28]\n" + "ld1b { z29.b }, p2/Z, [x24, x27]\n" + "ld1b { z28.b }, p2/Z, [x23, x27]\n" + "ld1b { z21.b }, p2/Z, [x22, x27]\n" + "ld1b { z27.b }, p2/Z, [x21, x27]\n" + "ld1b { z26.b }, p1/Z, [x24, x26]\n" + "ld1b { z25.b }, p1/Z, [x23, x26]\n" + "ld1b { z20.b }, p1/Z, [x22, x26]\n" + "ld1b { z24.b }, p1/Z, [x21, x26]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z4\n smax z19.b, p0/M, z19.b, z3.b\n" "movprfx z23, z2\n smax z23.b, p0/M, z23.b, z1.b\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" "movprfx z18, z0\n smax z18.b, p0/M, z18.b, z31.b\n" "smax z22.b, p0/M, z22.b, z30.b\n" - "ld1b { z4.b }, p4/Z, [x23, x28]\n" - "ld1b { z3.b }, p4/Z, [x22, x28]\n" + "ld1b { z4.b }, p4/Z, [x24, x9]\n" + "ld1b { z3.b }, p4/Z, [x23, x9]\n" "movprfx z17, z29\n smax z17.b, p0/M, z17.b, z28.b\n" "smax z21.b, p0/M, z21.b, z27.b\n" - "ld1b { z2.b }, p4/Z, [x21, x28]\n" - "ld1b { z1.b }, p4/Z, [x20, x28]\n" + "ld1b { z2.b }, p4/Z, [x22, x9]\n" + "ld1b { z1.b }, p4/Z, [x21, x9]\n" "movprfx z16, z26\n smax z16.b, p0/M, z16.b, z25.b\n" "smax z20.b, p0/M, z20.b, z24.b\n" - "ld1b { z0.b }, p3/Z, [x23, x27]\n" - "ld1b { z31.b }, p3/Z, [x22, x27]\n" + "ld1b { z0.b }, p3/Z, [x24, x28]\n" + "ld1b { z31.b }, p3/Z, [x23, x28]\n" "smax z19.b, p0/M, z19.b, z23.b\n" "smax z18.b, p0/M, z18.b, z22.b\n" - "ld1b { z22.b }, p3/Z, [x21, x27]\n" - "ld1b { z30.b }, p3/Z, [x20, x27]\n" + "ld1b { z22.b }, p3/Z, [x22, x28]\n" + "ld1b { z30.b }, p3/Z, [x21, x28]\n" "smax z17.b, p0/M, z17.b, z21.b\n" "smax z16.b, p0/M, z16.b, z20.b\n" - "ld1b { z29.b }, p2/Z, [x23, x26]\n" - "ld1b { z28.b }, p2/Z, [x22, x26]\n" - "subs x24, x24, #0x1\n" + "ld1b { z29.b }, p2/Z, [x24, x27]\n" + "ld1b { z28.b }, p2/Z, [x23, x27]\n" + "subs x25, x25, #0x1\n" "smax z8.b, p0/M, z8.b, z19.b\n" - "ld1b { z21.b }, p2/Z, [x21, x26]\n" - "ld1b { z27.b }, p2/Z, [x20, x26]\n" + "ld1b { z21.b }, p2/Z, [x22, x27]\n" + "ld1b { z27.b }, p2/Z, [x21, x27]\n" "smax z7.b, p0/M, z7.b, z18.b\n" "smax z6.b, p0/M, z6.b, z17.b\n" - "ld1b { z26.b }, p1/Z, [x23, x25]\n" - "ld1b { z25.b }, p1/Z, [x22, x25]\n" + "ld1b { z26.b }, p1/Z, [x24, x26]\n" + "ld1b { z25.b }, p1/Z, [x23, x26]\n" "smax z5.b, p0/M, z5.b, z16.b\n" - "add x19, x19, #0x20\n" - "ld1b { z20.b }, p1/Z, [x21, x25]\n" - "ld1b { z24.b }, p1/Z, [x20, x25]\n" + "add x20, x20, #0x20\n" + "ld1b { z20.b }, p1/Z, [x22, x26]\n" + "ld1b { z24.b }, p1/Z, [x21, x26]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "movprfx z19, z4\n smax z19.b, p0/M, z19.b, z3.b\n" @@ -137,33 +137,33 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smax z6.b, p0/M, z6.b, z17.b\n" "smax z5.b, p0/M, z5.b, z16.b\n" "4:" // 4-vectors of channels: After loop - "ands x20, %x[n_valid_cells], #0x3\n" + "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x23, [x19], #0x8\n" - "ld1b { z4.b }, p4/Z, [x23, x28]\n" - "subs x20, x20, #0x1\n" + "ldr x24, [x20], #0x8\n" + "ld1b { z4.b }, p4/Z, [x24, x9]\n" + "subs x21, x21, #0x1\n" "smax z8.b, p0/M, z8.b, z4.b\n" - "ld1b { z0.b }, p3/Z, [x23, x27]\n" - "ld1b { z29.b }, p2/Z, [x23, x26]\n" + "ld1b { z0.b }, p3/Z, [x24, x28]\n" + "ld1b { z29.b }, p2/Z, [x24, x27]\n" "smax z7.b, p0/M, z7.b, z0.b\n" "smax z6.b, p0/M, z6.b, z29.b\n" - "ld1b { z26.b }, p1/Z, [x23, x25]\n" + "ld1b { z26.b }, p1/Z, [x24, x26]\n" "smax z5.b, p0/M, z5.b, z26.b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End ".inst 0x4508a111 // sshllb z17.h, z8.b, #0x0\n" ".inst 0x4508a517 // sshllt z23.h, z8.b, #0x0\n" - "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1rw { z4.s }, p0/Z, [x19]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z4.s }, p0/Z, [x20]\n" ".inst 0x4508a0f6 // sshllb z22.h, z7.b, #0x0\n" ".inst 0x4508a4f5 // sshllt z21.h, z7.b, #0x0\n" - "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "ld1rw { z3.s }, p0/Z, [x19]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "ld1rw { z3.s }, p0/Z, [x20]\n" ".inst 0x4508a0d4 // sshllb z20.h, z6.b, #0x0\n" ".inst 0x4508a4d3 // sshllt z19.h, z6.b, #0x0\n" - "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1rw { z2.s }, p0/Z, [x19]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1rw { z2.s }, p0/Z, [x20]\n" ".inst 0x4508a0b2 // sshllb z18.h, z5.b, #0x0\n" ".inst 0x4508a4b0 // sshllt z16.h, z5.b, #0x0\n" ".inst 0x4510a221 // sshllb z1.s, z17.h, #0x0\n" @@ -259,7 +259,7 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smin z29.s, p0/M, z29.s, z18.s\n" "smin z28.s, p0/M, z28.s, z18.s\n" "trn1 z17.h, z30.h, z29.h\n" - "st1b { z16.b }, p4, [%x[outptr], x28]\n" + "st1b { z16.b }, p4, [%x[outptr], x9]\n" "smin z27.s, p0/M, z27.s, z18.s\n" "smin z26.s, p0/M, z26.s, z18.s\n" "trn1 z16.h, z28.h, z27.h\n" @@ -267,7 +267,7 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smin z25.s, p0/M, z25.s, z18.s\n" "smin z24.s, p0/M, z24.s, z18.s\n" "trn1 z17.h, z26.h, z25.h\n" - "st1b { z16.b }, p3, [%x[outptr], x27]\n" + "st1b { z16.b }, p3, [%x[outptr], x28]\n" "smin z23.s, p0/M, z23.s, z18.s\n" "smin z22.s, p0/M, z22.s, z18.s\n" "trn1 z16.h, z24.h, z23.h\n" @@ -275,47 +275,47 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smin z21.s, p0/M, z21.s, z18.s\n" "smin z20.s, p0/M, z20.s, z18.s\n" "trn1 z17.h, z22.h, z21.h\n" - "st1b { z16.b }, p2, [%x[outptr], x26]\n" + "st1b { z16.b }, p2, [%x[outptr], x27]\n" "smin z19.s, p0/M, z19.s, z18.s\n" "trn1 z16.h, z20.h, z19.h\n" "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p1, [%x[outptr], x25]\n" - "incb x25, ALL, MUL #4\n" - "whilelt p1.b, x25, %x[n_channels]\n" + "st1b { z16.b }, p1, [%x[outptr], x26]\n" + "incb x26, ALL, MUL #4\n" + "whilelt p1.b, x26, %x[n_channels]\n" + "incb x9, ALL, MUL #4\n" "incb x28, ALL, MUL #4\n" "incb x27, ALL, MUL #4\n" - "incb x26, ALL, MUL #4\n" "b.any 1b\n" "7:" // Single vector of channels - "whilelt p4.b, x28, %x[n_channels]\n" + "whilelt p4.b, x9, %x[n_channels]\n" "b.none 14f\n" "8:" // Single vector of channels: Loop - "lsr x24, %x[n_valid_cells], #0x2\n" + "lsr x25, %x[n_valid_cells], #0x2\n" "mov z8.b, #0x80\n" - "mov x19, %x[inptrs]\n" - "cbz x24, 11f\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" - "subs x24, x24, #0x1\n" - "add x19, x19, #0x20\n" - "ld1b { z4.b }, p4/Z, [x23, x28]\n" - "ld1b { z3.b }, p4/Z, [x22, x28]\n" - "ld1b { z2.b }, p4/Z, [x21, x28]\n" - "ld1b { z1.b }, p4/Z, [x20, x28]\n" + "mov x20, %x[inptrs]\n" + "cbz x25, 11f\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" + "subs x25, x25, #0x1\n" + "add x20, x20, #0x20\n" + "ld1b { z4.b }, p4/Z, [x24, x9]\n" + "ld1b { z3.b }, p4/Z, [x23, x9]\n" + "ld1b { z2.b }, p4/Z, [x22, x9]\n" + "ld1b { z1.b }, p4/Z, [x21, x9]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop "movprfx z19, z4\n smax z19.b, p0/M, z19.b, z3.b\n" "movprfx z23, z2\n smax z23.b, p0/M, z23.b, z1.b\n" - "ldp x23, x22, [x19, #0x0]\n" - "ldp x21, x20, [x19, #0x10]\n" + "ldp x24, x23, [x20, #0x0]\n" + "ldp x22, x21, [x20, #0x10]\n" "smax z19.b, p0/M, z19.b, z23.b\n" - "subs x24, x24, #0x1\n" - "ld1b { z4.b }, p4/Z, [x23, x28]\n" - "ld1b { z3.b }, p4/Z, [x22, x28]\n" + "subs x25, x25, #0x1\n" + "ld1b { z4.b }, p4/Z, [x24, x9]\n" + "ld1b { z3.b }, p4/Z, [x23, x9]\n" "smax z8.b, p0/M, z8.b, z19.b\n" - "add x19, x19, #0x20\n" - "ld1b { z2.b }, p4/Z, [x21, x28]\n" - "ld1b { z1.b }, p4/Z, [x20, x28]\n" + "add x20, x20, #0x20\n" + "ld1b { z2.b }, p4/Z, [x22, x9]\n" + "ld1b { z1.b }, p4/Z, [x21, x9]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail "movprfx z19, z4\n smax z19.b, p0/M, z19.b, z3.b\n" @@ -323,23 +323,23 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smax z19.b, p0/M, z19.b, z23.b\n" "smax z8.b, p0/M, z8.b, z19.b\n" "11:" // Single vector of channels: Loop: After loop - "ands x20, %x[n_valid_cells], #0x3\n" + "ands x21, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x23, [x19], #0x8\n" - "ld1b { z4.b }, p4/Z, [x23, x28]\n" - "subs x20, x20, #0x1\n" + "ldr x24, [x20], #0x8\n" + "ld1b { z4.b }, p4/Z, [x24, x9]\n" + "subs x21, x21, #0x1\n" "smax z8.b, p0/M, z8.b, z4.b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End ".inst 0x4508a111 // sshllb z17.h, z8.b, #0x0\n" ".inst 0x4508a517 // sshllt z23.h, z8.b, #0x0\n" - "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1rw { z4.s }, p0/Z, [x19]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z4.s }, p0/Z, [x20]\n" ".inst 0x4510a221 // sshllb z1.s, z17.h, #0x0\n" ".inst 0x4510a631 // sshllt z17.s, z17.h, #0x0\n" - "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "ld1rw { z3.s }, p0/Z, [x19]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "ld1rw { z3.s }, p0/Z, [x20]\n" ".inst 0x4510a2e0 // sshllb z0.s, z23.h, #0x0\n" ".inst 0x4510a6ff // sshllt z31.s, z23.h, #0x0\n" ".inst 0x44828081 // srshl z1.s, p0/M, z1.s, z4.s\n" @@ -348,8 +348,8 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( ".inst 0x4482809f // srshl z31.s, p0/M, z31.s, z4.s\n" ".inst 0x04a37421 // sqrdmulh z1.s, z1.s, z3.s\n" ".inst 0x04a37631 // sqrdmulh z17.s, z17.s, z3.s\n" - "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1rw { z2.s }, p0/Z, [x19]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1rw { z2.s }, p0/Z, [x20]\n" ".inst 0x04a37400 // sqrdmulh z0.s, z0.s, z3.s\n" ".inst 0x04a377ff // sqrdmulh z31.s, z31.s, z3.s\n" "mov z18.s, #0x7f\n" @@ -369,14 +369,14 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smin z31.s, p0/M, z31.s, z18.s\n" "trn1 z16.h, z0.h, z31.h\n" "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p4, [%x[outptr], x28]\n" - "incb x28\n" - "whilelt p4.b, x28, %x[n_channels]\n" + "st1b { z16.b }, p4, [%x[outptr], x9]\n" + "incb x9\n" + "whilelt p4.b, x9, %x[n_channels]\n" "b.any 8b\n" "14:" // End : : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [offsetof_qp_per_layer_left_shift] "I" (offsetof(Requantize32, per_layer_left_shift)), [offsetof_qp_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_qp_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [outptr] "r" (outptr), [quant_params] "r" (&qp) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } |