diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp | 221 |
1 files changed, 111 insertions, 110 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp index 3850ebf464..e9b586f4ce 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -22,10 +22,11 @@ * SOFTWARE. */ -#if defined(ARM_COMPUTE_ENABLE_SME) - -#include "src/core/NEON/kernels/assembly/pooling.hpp" +#include "pooling.hpp" #include <cstdint> +#include <cstddef> + +#if defined(ARM_COMPUTE_ENABLE_SME) namespace arm_conv { namespace pooling { @@ -56,68 +57,68 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "lsr x25, %x[n_valid_cells], #0x2\n" "mov z4.b, #0x80\n" "mov z3.b, #0x80\n" - "mov x20, %x[inptrs]\n" + "mov x24, %x[inptrs]\n" "mov z2.b, #0x80\n" "mov z1.b, #0x80\n" "cbz x25, 4f\n" - "ldp x24, x23, [x20, #0x0]\n" + "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "ld1b { z0.b }, p4/Z, [x24, x9]\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" - "ld1b { z31.b }, p4/Z, [x23, x9]\n" - "ld1b { z23.b }, p4/Z, [x22, x9]\n" - "ld1b { z30.b }, p4/Z, [x21, x9]\n" - "ld1b { z18.b }, p3/Z, [x24, x28]\n" - "ld1b { z29.b }, p3/Z, [x23, x28]\n" - "ld1b { z22.b }, p3/Z, [x22, x28]\n" - "ld1b { z28.b }, p3/Z, [x21, x28]\n" - "ld1b { z17.b }, p2/Z, [x24, x27]\n" - "ld1b { z27.b }, p2/Z, [x23, x27]\n" - "ld1b { z21.b }, p2/Z, [x22, x27]\n" - "ld1b { z26.b }, p2/Z, [x21, x27]\n" - "ld1b { z16.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" + "ld1b { z0.b }, p4/Z, [x23, x9]\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" + "ld1b { z18.b }, p3/Z, [x23, x28]\n" + "ld1b { z29.b }, p3/Z, [x22, x28]\n" + "ld1b { z22.b }, p3/Z, [x21, x28]\n" + "ld1b { z28.b }, p3/Z, [x20, x28]\n" + "ld1b { z17.b }, p2/Z, [x23, x27]\n" + "ld1b { z27.b }, p2/Z, [x22, x27]\n" + "ld1b { z21.b }, p2/Z, [x21, x27]\n" + "ld1b { z26.b }, p2/Z, [x20, x27]\n" + "ld1b { z16.b }, p1/Z, [x23, x26]\n" + "ld1b { z25.b }, p1/Z, [x22, x26]\n" + "ld1b { z20.b }, p1/Z, [x21, x26]\n" + "ld1b { z24.b }, p1/Z, [x20, x26]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z0\n smax z19.b, p0/M, z19.b, z31.b\n" "smax z23.b, p0/M, z23.b, z30.b\n" - "ldp x24, x23, [x20, #0x0]\n" + "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" "smax z18.b, p0/M, z18.b, z29.b\n" "smax z22.b, p0/M, z22.b, z28.b\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" "smax z17.b, p0/M, z17.b, z27.b\n" "smax z21.b, p0/M, z21.b, z26.b\n" - "ld1b { z0.b }, p4/Z, [x24, x9]\n" + "ld1b { z0.b }, p4/Z, [x23, x9]\n" "smax z16.b, p0/M, z16.b, z25.b\n" "smax z20.b, p0/M, z20.b, z24.b\n" - "ld1b { z31.b }, p4/Z, [x23, x9]\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" "smax z19.b, p0/M, z19.b, z23.b\n" "smax z18.b, p0/M, z18.b, z22.b\n" - "ld1b { z23.b }, p4/Z, [x22, x9]\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" "smax z17.b, p0/M, z17.b, z21.b\n" "smax z16.b, p0/M, z16.b, z20.b\n" - "ld1b { z30.b }, p4/Z, [x21, x9]\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" "smax z4.b, p0/M, z4.b, z19.b\n" "smax z3.b, p0/M, z3.b, z18.b\n" - "ld1b { z18.b }, p3/Z, [x24, x28]\n" + "ld1b { z18.b }, p3/Z, [x23, x28]\n" "smax z2.b, p0/M, z2.b, z17.b\n" "smax z1.b, p0/M, z1.b, z16.b\n" - "ld1b { z29.b }, p3/Z, [x23, x28]\n" - "ld1b { z22.b }, p3/Z, [x22, x28]\n" - "ld1b { z28.b }, p3/Z, [x21, x28]\n" - "ld1b { z17.b }, p2/Z, [x24, x27]\n" - "ld1b { z27.b }, p2/Z, [x23, x27]\n" - "ld1b { z21.b }, p2/Z, [x22, x27]\n" - "ld1b { z26.b }, p2/Z, [x21, x27]\n" - "ld1b { z16.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" + "ld1b { z29.b }, p3/Z, [x22, x28]\n" + "ld1b { z22.b }, p3/Z, [x21, x28]\n" + "ld1b { z28.b }, p3/Z, [x20, x28]\n" + "ld1b { z17.b }, p2/Z, [x23, x27]\n" + "ld1b { z27.b }, p2/Z, [x22, x27]\n" + "ld1b { z21.b }, p2/Z, [x21, x27]\n" + "ld1b { z26.b }, p2/Z, [x20, x27]\n" + "ld1b { z16.b }, p1/Z, [x23, x26]\n" + "ld1b { z25.b }, p1/Z, [x22, x26]\n" + "ld1b { z20.b }, p1/Z, [x21, x26]\n" + "ld1b { z24.b }, p1/Z, [x20, x26]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "movprfx z19, z0\n smax z19.b, p0/M, z19.b, z31.b\n" @@ -140,15 +141,15 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x24, [x20], #0x8\n" - "ld1b { z0.b }, p4/Z, [x24, x9]\n" + "ldr x20, [x24], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x9]\n" "subs x21, x21, #0x1\n" - "smax z4.b, p0/M, z4.b, z0.b\n" - "ld1b { z18.b }, p3/Z, [x24, x28]\n" - "smax z3.b, p0/M, z3.b, z18.b\n" - "ld1b { z17.b }, p2/Z, [x24, x27]\n" - "smax z2.b, p0/M, z2.b, z17.b\n" - "ld1b { z16.b }, p1/Z, [x24, x26]\n" + "smax z4.b, p0/M, z4.b, z16.b\n" + "ld1b { z16.b }, p3/Z, [x20, x28]\n" + "smax z3.b, p0/M, z3.b, z16.b\n" + "ld1b { z16.b }, p2/Z, [x20, x27]\n" + "smax z2.b, p0/M, z2.b, z16.b\n" + "ld1b { z16.b }, p1/Z, [x20, x26]\n" "smax z1.b, p0/M, z1.b, z16.b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End @@ -292,83 +293,83 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" "mov z4.b, #0x80\n" - "mov x20, %x[inptrs]\n" + "mov x24, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x24, x23, [x20, #0x0]\n" + "ldp x20, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "ld1b { z0.b }, p4/Z, [x24, x9]\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" - "ld1b { z31.b }, p4/Z, [x23, x9]\n" - "ld1b { z23.b }, p4/Z, [x22, x9]\n" - "ld1b { z30.b }, p4/Z, [x21, x9]\n" + "ld1b { z0.b }, p4/Z, [x20, x9]\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop - "movprfx z19, z0\n smax z19.b, p0/M, z19.b, z31.b\n" - "smax z23.b, p0/M, z23.b, z30.b\n" - "ldp x24, x23, [x20, #0x0]\n" + "movprfx z16, z0\n smax z16.b, p0/M, z16.b, z31.b\n" + "movprfx z17, z23\n smax z17.b, p0/M, z17.b, z30.b\n" + "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "smax z19.b, p0/M, z19.b, z23.b\n" - "ldp x22, x21, [x20, #0x10]\n" - "smax z4.b, p0/M, z4.b, z19.b\n" - "add x20, x20, #0x20\n" - "ld1b { z0.b }, p4/Z, [x24, x9]\n" - "ld1b { z31.b }, p4/Z, [x23, x9]\n" - "ld1b { z23.b }, p4/Z, [x22, x9]\n" - "ld1b { z30.b }, p4/Z, [x21, x9]\n" + "smax z16.b, p0/M, z16.b, z17.b\n" + "ldp x21, x20, [x24, #0x10]\n" + "smax z4.b, p0/M, z4.b, z16.b\n" + "add x24, x24, #0x20\n" + "ld1b { z0.b }, p4/Z, [x23, x9]\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail - "movprfx z19, z0\n smax z19.b, p0/M, z19.b, z31.b\n" - "smax z23.b, p0/M, z23.b, z30.b\n" - "smax z19.b, p0/M, z19.b, z23.b\n" - "smax z4.b, p0/M, z4.b, z19.b\n" + "movprfx z16, z0\n smax z16.b, p0/M, z16.b, z31.b\n" + "movprfx z17, z23\n smax z17.b, p0/M, z17.b, z30.b\n" + "smax z16.b, p0/M, z16.b, z17.b\n" + "smax z4.b, p0/M, z4.b, z16.b\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 x24, [x20], #0x8\n" - "ld1b { z0.b }, p4/Z, [x24, x9]\n" + "ldr x20, [x24], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x9]\n" "subs x21, x21, #0x1\n" - "smax z4.b, p0/M, z4.b, z0.b\n" + "smax z4.b, p0/M, z4.b, z16.b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - ".inst 0x4508a097 // sshllb z23.h, z4.b, #0x0\n" - ".inst 0x4508a496 // sshllt z22.h, z4.b, #0x0\n" + ".inst 0x4508a091 // sshllb z17.h, z4.b, #0x0\n" + ".inst 0x4508a490 // sshllt z16.h, z4.b, #0x0\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1rw { z4.s }, p0/Z, [x20]\n" - ".inst 0x4510a2e1 // sshllb z1.s, z23.h, #0x0\n" - ".inst 0x4510a6f7 // sshllt z23.s, z23.h, #0x0\n" + "ld1rw { z18.s }, p0/Z, [x20]\n" + ".inst 0x4510a236 // sshllb z22.s, z17.h, #0x0\n" + ".inst 0x4510a635 // sshllt z21.s, z17.h, #0x0\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "ld1rw { z3.s }, p0/Z, [x20]\n" - ".inst 0x4510a2c0 // sshllb z0.s, z22.h, #0x0\n" - ".inst 0x4510a6df // sshllt z31.s, z22.h, #0x0\n" + "ld1rw { z17.s }, p0/Z, [x20]\n" + ".inst 0x4510a214 // sshllb z20.s, z16.h, #0x0\n" + ".inst 0x4510a613 // sshllt z19.s, z16.h, #0x0\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1rw { z2.s }, p0/Z, [x20]\n" - ".inst 0x44828081 // srshl z1.s, p0/M, z1.s, z4.s\n" - ".inst 0x44828097 // srshl z23.s, p0/M, z23.s, z4.s\n" - ".inst 0x44828080 // srshl z0.s, p0/M, z0.s, z4.s\n" - ".inst 0x4482809f // srshl z31.s, p0/M, z31.s, z4.s\n" - ".inst 0x04a37421 // sqrdmulh z1.s, z1.s, z3.s\n" - ".inst 0x04a376f7 // sqrdmulh z23.s, z23.s, z3.s\n" - ".inst 0x04a37400 // sqrdmulh z0.s, z0.s, z3.s\n" - ".inst 0x04a377ff // sqrdmulh z31.s, z31.s, z3.s\n" - "mov z19.s, #0x7f\n" - ".inst 0x44828041 // srshl z1.s, p0/M, z1.s, z2.s\n" - ".inst 0x44828057 // srshl z23.s, p0/M, z23.s, z2.s\n" - ".inst 0x44828040 // srshl z0.s, p0/M, z0.s, z2.s\n" - ".inst 0x4482805f // srshl z31.s, p0/M, z31.s, z2.s\n" - "not z16.s, p0/M, z19.s\n" - "smax z1.s, p0/M, z1.s, z16.s\n" - "smax z23.s, p0/M, z23.s, z16.s\n" - "smax z0.s, p0/M, z0.s, z16.s\n" - "smax z31.s, p0/M, z31.s, z16.s\n" - "smin z1.s, p0/M, z1.s, z19.s\n" - "smin z23.s, p0/M, z23.s, z19.s\n" - "trn1 z23.h, z1.h, z23.h\n" - "smin z0.s, p0/M, z0.s, z19.s\n" - "smin z31.s, p0/M, z31.s, z19.s\n" - "trn1 z16.h, z0.h, z31.h\n" - "trn1 z16.b, z23.b, z16.b\n" + "ld1rw { z16.s }, p0/Z, [x20]\n" + ".inst 0x44828256 // srshl z22.s, p0/M, z22.s, z18.s\n" + ".inst 0x44828255 // srshl z21.s, p0/M, z21.s, z18.s\n" + ".inst 0x44828254 // srshl z20.s, p0/M, z20.s, z18.s\n" + ".inst 0x44828253 // srshl z19.s, p0/M, z19.s, z18.s\n" + ".inst 0x04b176d6 // sqrdmulh z22.s, z22.s, z17.s\n" + ".inst 0x04b176b5 // sqrdmulh z21.s, z21.s, z17.s\n" + ".inst 0x04b17694 // sqrdmulh z20.s, z20.s, z17.s\n" + ".inst 0x04b17673 // sqrdmulh z19.s, z19.s, z17.s\n" + "mov z18.s, #0x7f\n" + ".inst 0x44828216 // srshl z22.s, p0/M, z22.s, z16.s\n" + ".inst 0x44828215 // srshl z21.s, p0/M, z21.s, z16.s\n" + ".inst 0x44828214 // srshl z20.s, p0/M, z20.s, z16.s\n" + ".inst 0x44828213 // srshl z19.s, p0/M, z19.s, z16.s\n" + "not z16.s, p0/M, z18.s\n" + "smax z22.s, p0/M, z22.s, z16.s\n" + "smax z21.s, p0/M, z21.s, z16.s\n" + "smax z20.s, p0/M, z20.s, z16.s\n" + "smax z19.s, p0/M, z19.s, z16.s\n" + "smin z22.s, p0/M, z22.s, z18.s\n" + "smin z21.s, p0/M, z21.s, z18.s\n" + "trn1 z17.h, z22.h, z21.h\n" + "smin z20.s, p0/M, z20.s, z18.s\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 }, p4, [%x[outptr], x9]\n" "incb x9\n" "whilelt p4.b, x9, %x[n_channels]\n" |