diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp | 148 |
1 files changed, 74 insertions, 74 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp index 7d14edddeb..da9e1408f9 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8_nhwc_max_generic_depthfirst/generic.cpp @@ -54,68 +54,68 @@ void sve_s8_nhwc_max_generic_depthfirst_impl( "lsr x25, %x[n_valid_cells], #0x2\n" "mov z8.b, #0x80\n" "mov z7.b, #0x80\n" - "mov x20, %x[inptrs]\n" + "mov x24, %x[inptrs]\n" "mov z6.b, #0x80\n" "mov z5.b, #0x80\n" "cbz x25, 4f\n" - "ldp x24, x23, [x20, #0x0]\n" - "ldp x22, x21, [x20, #0x10]\n" + "ldp x23, x22, [x24, #0x0]\n" + "ldp x21, x20, [x24, #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" + "add x24, x24, #0x20\n" + "ld1b { z4.b }, p4/Z, [x23, x9]\n" + "ld1b { z3.b }, p4/Z, [x22, x9]\n" + "ld1b { z2.b }, p4/Z, [x21, x9]\n" + "ld1b { z1.b }, p4/Z, [x20, x9]\n" + "ld1b { z0.b }, p3/Z, [x23, x28]\n" + "ld1b { z31.b }, p3/Z, [x22, x28]\n" + "ld1b { z22.b }, p3/Z, [x21, x28]\n" + "ld1b { z30.b }, p3/Z, [x20, x28]\n" + "ld1b { z29.b }, p2/Z, [x23, x27]\n" + "ld1b { z28.b }, p2/Z, [x22, x27]\n" + "ld1b { z21.b }, p2/Z, [x21, x27]\n" + "ld1b { z27.b }, p2/Z, [x20, x27]\n" + "ld1b { z26.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, 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 x24, x23, [x20, #0x0]\n" - "ldp x22, x21, [x20, #0x10]\n" + "ldp x23, x22, [x24, #0x0]\n" + "ldp x21, x20, [x24, #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, [x24, x9]\n" - "ld1b { z3.b }, p4/Z, [x23, x9]\n" + "ld1b { z4.b }, p4/Z, [x23, x9]\n" + "ld1b { z3.b }, p4/Z, [x22, 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, [x22, x9]\n" - "ld1b { z1.b }, p4/Z, [x21, x9]\n" + "ld1b { z2.b }, p4/Z, [x21, x9]\n" + "ld1b { z1.b }, p4/Z, [x20, 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, [x24, x28]\n" - "ld1b { z31.b }, p3/Z, [x23, x28]\n" + "ld1b { z0.b }, p3/Z, [x23, x28]\n" + "ld1b { z31.b }, p3/Z, [x22, 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, [x22, x28]\n" - "ld1b { z30.b }, p3/Z, [x21, x28]\n" + "ld1b { z22.b }, p3/Z, [x21, x28]\n" + "ld1b { z30.b }, p3/Z, [x20, 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, [x24, x27]\n" - "ld1b { z28.b }, p2/Z, [x23, x27]\n" + "ld1b { z29.b }, p2/Z, [x23, x27]\n" + "ld1b { z28.b }, p2/Z, [x22, x27]\n" "subs x25, x25, #0x1\n" "smax z8.b, p0/M, z8.b, z19.b\n" - "ld1b { z21.b }, p2/Z, [x22, x27]\n" - "ld1b { z27.b }, p2/Z, [x21, x27]\n" + "ld1b { z21.b }, p2/Z, [x21, x27]\n" + "ld1b { z27.b }, p2/Z, [x20, 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, [x24, x26]\n" - "ld1b { z25.b }, p1/Z, [x23, x26]\n" + "ld1b { z26.b }, p1/Z, [x23, x26]\n" + "ld1b { z25.b }, p1/Z, [x22, x26]\n" "smax z5.b, p0/M, z5.b, z16.b\n" - "add x20, x20, #0x20\n" - "ld1b { z20.b }, p1/Z, [x22, x26]\n" - "ld1b { z24.b }, p1/Z, [x21, x26]\n" + "add x24, x24, #0x20\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, z4\n smax z19.b, p0/M, z19.b, z3.b\n" @@ -138,16 +138,16 @@ void sve_s8_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 { z4.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 z8.b, p0/M, z8.b, z4.b\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, [x24, x26]\n" - "smax z5.b, p0/M, z5.b, z26.b\n" + "smax z8.b, p0/M, z8.b, z16.b\n" + "ld1b { z17.b }, p3/Z, [x20, x28]\n" + "ld1b { z16.b }, p2/Z, [x20, x27]\n" + "smax z7.b, p0/M, z7.b, z17.b\n" + "smax z6.b, p0/M, z6.b, z16.b\n" + "ld1b { z16.b }, p1/Z, [x20, x26]\n" + "smax z5.b, p0/M, z5.b, z16.b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End "st1b { z8.b }, p4, [%x[outptr], x9]\n" @@ -166,44 +166,44 @@ void sve_s8_nhwc_max_generic_depthfirst_impl( "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" "mov z8.b, #0x80\n" - "mov x20, %x[inptrs]\n" + "mov x24, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x24, x23, [x20, #0x0]\n" - "ldp x22, x21, [x20, #0x10]\n" + "ldp x23, x22, [x24, #0x0]\n" + "ldp x21, x20, [x24, #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" + "add x24, x24, #0x20\n" + "ld1b { z4.b }, p4/Z, [x23, x9]\n" + "ld1b { z3.b }, p4/Z, [x22, x9]\n" + "ld1b { z2.b }, p4/Z, [x21, x9]\n" + "ld1b { z1.b }, p4/Z, [x20, 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 x24, x23, [x20, #0x0]\n" - "ldp x22, x21, [x20, #0x10]\n" - "smax z19.b, p0/M, z19.b, z23.b\n" + "movprfx z16, z4\n smax z16.b, p0/M, z16.b, z3.b\n" + "movprfx z17, z2\n smax z17.b, p0/M, z17.b, z1.b\n" + "ldp x23, x22, [x24, #0x0]\n" + "ldp x21, x20, [x24, #0x10]\n" + "smax z16.b, p0/M, z16.b, z17.b\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 x20, x20, #0x20\n" - "ld1b { z2.b }, p4/Z, [x22, x9]\n" - "ld1b { z1.b }, p4/Z, [x21, x9]\n" + "ld1b { z4.b }, p4/Z, [x23, x9]\n" + "ld1b { z3.b }, p4/Z, [x22, x9]\n" + "smax z8.b, p0/M, z8.b, z16.b\n" + "add x24, x24, #0x20\n" + "ld1b { z2.b }, p4/Z, [x21, x9]\n" + "ld1b { z1.b }, p4/Z, [x20, 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" - "movprfx z23, z2\n smax z23.b, p0/M, z23.b, z1.b\n" - "smax z19.b, p0/M, z19.b, z23.b\n" - "smax z8.b, p0/M, z8.b, z19.b\n" + "movprfx z16, z4\n smax z16.b, p0/M, z16.b, z3.b\n" + "movprfx z17, z2\n smax z17.b, p0/M, z17.b, z1.b\n" + "smax z16.b, p0/M, z16.b, z17.b\n" + "smax z8.b, p0/M, z8.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 { z4.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 z8.b, p0/M, z8.b, z4.b\n" + "smax z8.b, p0/M, z8.b, z16.b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End "st1b { z8.b }, p4, [%x[outptr], x9]\n" |