aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp454
1 files changed, 227 insertions, 227 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
index 0a4929918b..7c5f824acf 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
@@ -89,245 +89,245 @@ void sve_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
__asm__ __volatile__(
"ptrue p3.b\n"
"ldr x20, [%x[params_struct], %[offsetof_args_outptrs]]\n"
- "ldr x17, [%x[params_struct], %[offsetof_args_params]]\n"
- "add x16, %x[params_struct], %[offsetof_Args_inptrs]\n"
- "cntw x15\n"
- "mov x14, #0x0\n"
- "whilelt p2.s, XZR, %x[n_channels]\n"
+ "ldr x16, [%x[params_struct], %[offsetof_args_params]]\n"
+ "add x15, %x[params_struct], %[offsetof_Args_inptrs]\n"
+ "cntw x14\n"
"ldp x13, x12, [x20, #0x0]\n"
"ldp x11, x10, [x20, #0x10]\n"
- "ld1w { z30.s }, p3/Z, [x17]\n"
- "ld1w { z0.s }, p3/Z, [x17, #1, MUL VL]\n"
- "ld1w { z1.s }, p3/Z, [x17, #2, MUL VL]\n"
- "ld1w { z2.s }, p3/Z, [x17, #3, MUL VL]\n"
- "cmp x15, %x[n_channels]\n"
- "sub x9, XZR, x15\n"
- "ld1w { z3.s }, p3/Z, [x17, #4, MUL VL]\n"
- "ld1w { z4.s }, p3/Z, [x17, #5, MUL VL]\n"
- "ld1w { z5.s }, p3/Z, [x17, #6, MUL VL]\n"
- "ld1w { z6.s }, p3/Z, [x17, #7, MUL VL]\n"
- "addvl x17, x17, #16\n"
- "ldp x27, x26, [x16, #0x0]\n"
- "ldp x25, x24, [x16, #0x10]\n"
- "ldp x23, x22, [x16, #0x20]\n"
- "ldp x21, x20, [x16, #0x30]\n"
- "ld1rw { z29.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
- "ld1rw { z28.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
- "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n"
- "ld1w { z8.s }, p3/Z, [x17, #-7, MUL VL]\n"
- "addvl x17, x17, #-6\n"
- "ld1w { z9.s }, p2/Z, [x27, x14, LSL #2]\n"
- "ld1w { z10.s }, p2/Z, [x26, x14, LSL #2]\n"
- "ld1w { z11.s }, p2/Z, [x25, x14, LSL #2]\n"
- "ld1w { z12.s }, p2/Z, [x24, x14, LSL #2]\n"
- "ld1w { z13.s }, p2/Z, [x23, x14, LSL #2]\n"
- "ld1w { z14.s }, p2/Z, [x22, x14, LSL #2]\n"
- "ld1w { z15.s }, p2/Z, [x21, x14, LSL #2]\n"
- "ld1w { z16.s }, p2/Z, [x20, x14, LSL #2]\n"
+ "mov x9, #0x0\n"
+ "whilelt p2.s, XZR, %x[n_channels]\n"
+ "ld1w { z20.s }, p3/Z, [x16]\n"
+ "ld1w { z0.s }, p3/Z, [x16, #1, MUL VL]\n"
+ "cmp x14, %x[n_channels]\n"
+ "ld1w { z1.s }, p3/Z, [x16, #2, MUL VL]\n"
+ "ld1w { z2.s }, p3/Z, [x16, #3, MUL VL]\n"
+ "sub x28, XZR, x14\n"
+ "ld1w { z3.s }, p3/Z, [x16, #4, MUL VL]\n"
+ "ld1w { z4.s }, p3/Z, [x16, #5, MUL VL]\n"
+ "ld1w { z5.s }, p3/Z, [x16, #6, MUL VL]\n"
+ "ld1w { z6.s }, p3/Z, [x16, #7, MUL VL]\n"
+ "addvl x16, x16, #16\n"
+ "ldp x27, x26, [x15, #0x0]\n"
+ "ldp x25, x24, [x15, #0x10]\n"
+ "ldp x23, x22, [x15, #0x20]\n"
+ "ldp x21, x20, [x15, #0x30]\n"
+ "ld1rw { z26.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
+ "ld1rw { z25.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
+ "ld1w { z7.s }, p3/Z, [x16, #-8, MUL VL]\n"
+ "ld1w { z8.s }, p3/Z, [x16, #-7, MUL VL]\n"
+ "addvl x16, x16, #-6\n"
+ "ld1w { z9.s }, p2/Z, [x27, x9, LSL #2]\n"
+ "ld1w { z10.s }, p2/Z, [x26, x9, LSL #2]\n"
+ "ld1w { z11.s }, p2/Z, [x25, x9, LSL #2]\n"
+ "ld1w { z12.s }, p2/Z, [x24, x9, LSL #2]\n"
+ "ld1w { z13.s }, p2/Z, [x23, x9, LSL #2]\n"
+ "ld1w { z14.s }, p2/Z, [x22, x9, LSL #2]\n"
+ "ld1w { z15.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
"bge 2f\n"
"1:" // Channel loop
- "movprfx z27, z30\n fmla z27.s, p3/M, z8.s, z9.s\n"
- "movprfx z26, z30\n fmla z26.s, p3/M, z6.s, z9.s\n"
- "ldr x28, [x16, #0x40]\n"
- "ldr x21, [x16, #0x48]\n"
- "ldr x25, [x16, #0x50]\n"
- "ldr x20, [x16, #0x58]\n"
- "movprfx z25, z30\n fmla z25.s, p3/M, z2.s, z9.s\n"
- "movprfx z24, z30\n fmla z24.s, p3/M, z0.s, z9.s\n"
- "ldr x27, [x16, #0x78]\n"
- "ldr x24, [x16, #0x60]\n"
- "whilelt p1.s, x15, %x[n_channels]\n"
+ "movprfx z24, z20\n fmla z24.s, p3/M, z8.s, z9.s\n"
+ "movprfx z23, z20\n fmla z23.s, p3/M, z6.s, z9.s\n"
+ "ldr x21, [x15, #0x40]\n"
+ "ldr x20, [x15, #0x48]\n"
+ "fmla z24.s, p3/M, z0.s, z10.s\n"
+ "fmla z23.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x20, [x15, #0x50]\n"
+ "fmla z24.s, p3/M, z1.s, z11.s\n"
+ "fmla z23.s, p3/M, z2.s, z13.s\n"
+ "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z19.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z24.s, p3/M, z3.s, z14.s\n"
+ "fmla z23.s, p3/M, z0.s, z16.s\n"
+ "ldr x20, [x15, #0x58]\n"
+ "ldr x22, [x15, #0x78]\n"
+ "fmla z24.s, p3/M, z4.s, z15.s\n"
+ "fmla z23.s, p3/M, z4.s, z17.s\n"
+ "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x21, [x15, #0x60]\n"
+ "fmla z24.s, p3/M, z2.s, z16.s\n"
+ "fmla z23.s, p3/M, z5.s, z18.s\n"
+ "ldr x20, [x15, #0x80]\n"
+ "ld1w { z18.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "movprfx z22, z20\n fmla z22.s, p3/M, z2.s, z9.s\n"
+ "movprfx z21, z20\n fmla z21.s, p3/M, z0.s, z9.s\n"
+ "ld1w { z20.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x21, [x15, #0x68]\n"
+ "fmla z24.s, p3/M, z5.s, z19.s\n"
+ "fmla z23.s, p3/M, z3.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x22, x9, LSL #2]\n"
+ "ldr x20, [x15, #0x88]\n"
+ "fmla z22.s, p3/M, z3.s, z17.s\n"
+ "fmla z21.s, p3/M, z4.s, z16.s\n"
+ "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z0.s, z18.s\n"
+ "fmla z21.s, p3/M, z1.s, z20.s\n"
+ "ldr x21, [x15, #0x70]\n"
+ "ldr x20, [x15, #0x98]\n"
+ "fmla z22.s, p3/M, z4.s, z17.s\n"
+ "fmla z21.s, p3/M, z5.s, z16.s\n"
+ "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z19.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z24.s, p3/M, z6.s, z18.s\n"
+ "ldr x21, [x15, #0x90]\n"
+ "ldr x20, [x15, #0xa8]\n"
+ "fmla z22.s, p3/M, z1.s, z16.s\n"
+ "fmla z21.s, p3/M, z2.s, z19.s\n"
+ "fmla z24.s, p3/M, z7.s, z16.s\n"
+ "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x21, [x15, #0xa0]\n"
+ "ldr x20, [x15, #0xb0]\n"
+ "fmla z22.s, p3/M, z6.s, z16.s\n"
+ "fmla z21.s, p3/M, z3.s, z18.s\n"
+ "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z7.s, z17.s\n"
+ "fmla z21.s, p3/M, z7.s, z16.s\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "fmla z23.s, p3/M, z7.s, z20.s\n"
+ "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z5.s, z18.s\n"
+ "ldr x20, [x15, #0xc0]\n"
+ "fmla z21.s, p3/M, z6.s, z17.s\n"
+ "fmla z23.s, p3/M, z8.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z8.s, z17.s\n"
+ "fmla z21.s, p3/M, z8.s, z16.s\n"
+ "whilelt p1.s, x14, %x[n_channels]\n"
+ "ldp x27, x26, [x15, #0x0]\n"
+ "ldp x25, x24, [x15, #0x10]\n"
+ "ldp x23, x22, [x15, #0x20]\n"
"incw x9\n"
- "ldr x23, [x16, #0x68]\n"
- "ldr x26, [x16, #0x70]\n"
+ "fmax z24.s, p3/M, z24.s, z26.s\n"
+ "ldp x21, x20, [x15, #0x30]\n"
+ "ld1w { z9.s }, p1/Z, [x27, x14, LSL #2]\n"
+ "fmax z23.s, p3/M, z23.s, z26.s\n"
+ "fmax z22.s, p3/M, z22.s, z26.s\n"
+ "ld1w { z10.s }, p1/Z, [x26, x14, LSL #2]\n"
+ "ld1w { z11.s }, p1/Z, [x25, x14, LSL #2]\n"
+ "fmax z21.s, p3/M, z21.s, z26.s\n"
+ "incw x28\n"
+ "ld1w { z12.s }, p1/Z, [x24, x14, LSL #2]\n"
+ "ld1w { z13.s }, p1/Z, [x23, x14, LSL #2]\n"
"mov p0.b, p2.b\n"
- "fmla z27.s, p3/M, z0.s, z10.s\n"
- "fmla z26.s, p3/M, z1.s, z12.s\n"
- "ld1w { z21.s }, p2/Z, [x21, x14, LSL #2]\n"
- "ldr x22, [x16, #0x88]\n"
- "ld1w { z30.s }, p3/Z, [x17]\n"
- "fmla z27.s, p3/M, z1.s, z11.s\n"
- "ld1w { z18.s }, p2/Z, [x28, x14, LSL #2]\n"
- "ldr x21, [x16, #0x80]\n"
- "fmla z26.s, p3/M, z2.s, z13.s\n"
- "ld1w { z20.s }, p2/Z, [x25, x14, LSL #2]\n"
- "ldr x25, [x16, #0x90]\n"
- "fmla z27.s, p3/M, z3.s, z14.s\n"
- "ld1w { z17.s }, p2/Z, [x20, x14, LSL #2]\n"
- "ldr x20, [x16, #0x98]\n"
- "fmla z26.s, p3/M, z0.s, z16.s\n"
- "fmla z25.s, p3/M, z3.s, z17.s\n"
- "ld1w { z23.s }, p2/Z, [x22, x14, LSL #2]\n"
- "fmla z27.s, p3/M, z4.s, z15.s\n"
- "ld1w { z22.s }, p2/Z, [x24, x14, LSL #2]\n"
- "ldr x24, [x16, #0xa0]\n"
- "fmla z26.s, p3/M, z4.s, z18.s\n"
- "ld1w { z17.s }, p2/Z, [x23, x14, LSL #2]\n"
- "ldr x23, [x16, #0xa8]\n"
- "fmla z25.s, p3/M, z0.s, z22.s\n"
- "ld1w { z0.s }, p3/Z, [x17, #1, MUL VL]\n"
- "fmla z27.s, p3/M, z2.s, z16.s\n"
- "ld1w { z19.s }, p2/Z, [x26, x14, LSL #2]\n"
- "ldr x22, [x16, #0xb0]\n"
- "fmla z26.s, p3/M, z5.s, z21.s\n"
- "ld1w { z18.s }, p2/Z, [x21, x14, LSL #2]\n"
- "ldr x21, [x16, #0xc0]\n"
- "fmla z25.s, p3/M, z4.s, z17.s\n"
- "ld1w { z21.s }, p2/Z, [x20, x14, LSL #2]\n"
- "fmla z27.s, p3/M, z5.s, z20.s\n"
- "fmla z26.s, p3/M, z3.s, z20.s\n"
- "ld1w { z16.s }, p2/Z, [x27, x14, LSL #2]\n"
- "ldr x20, [x16, #0xb8]\n"
- "fmla z24.s, p3/M, z4.s, z16.s\n"
- "ld1w { z20.s }, p2/Z, [x24, x14, LSL #2]\n"
- "ld1w { z4.s }, p3/Z, [x17, #5, MUL VL]\n"
- "fmla z27.s, p3/M, z6.s, z22.s\n"
- "ld1w { z17.s }, p2/Z, [x25, x14, LSL #2]\n"
- "fmla z25.s, p3/M, z1.s, z19.s\n"
- "fmla z24.s, p3/M, z1.s, z18.s\n"
- "ld1w { z1.s }, p3/Z, [x17, #2, MUL VL]\n"
- "fmla z26.s, p3/M, z7.s, z18.s\n"
- "fmla z27.s, p3/M, z7.s, z19.s\n"
- "ld1w { z16.s }, p2/Z, [x23, x14, LSL #2]\n"
- "fmla z24.s, p3/M, z5.s, z23.s\n"
- "ld1w { z19.s }, p2/Z, [x22, x14, LSL #2]\n"
- "fmla z25.s, p3/M, z6.s, z17.s\n"
- "ld1w { z18.s }, p2/Z, [x20, x14, LSL #2]\n"
- "fmla z26.s, p3/M, z8.s, z21.s\n"
- "fmax z27.s, p3/M, z27.s, z29.s\n"
- "fmla z24.s, p3/M, z2.s, z21.s\n"
- "ld1w { z17.s }, p2/Z, [x21, x14, LSL #2]\n"
- "ldp x27, x26, [x16, #0x0]\n"
- "ldp x25, x24, [x16, #0x10]\n"
+ "whilelt p2.s, x9, %x[n_channels]\n"
+ "ld1w { z14.s }, p1/Z, [x22, x14, LSL #2]\n"
+ "ld1w { z15.s }, p1/Z, [x21, x14, LSL #2]\n"
+ "fmin z24.s, p3/M, z24.s, z25.s\n"
+ "fmin z23.s, p3/M, z23.s, z25.s\n"
+ "ld1w { z16.s }, p1/Z, [x20, x14, LSL #2]\n"
"incw x14\n"
- "ldp x23, x22, [x16, #0x20]\n"
- "ldp x21, x20, [x16, #0x30]\n"
- "fmla z25.s, p3/M, z7.s, z20.s\n"
- "ld1w { z2.s }, p3/Z, [x17, #3, MUL VL]\n"
- "fmin z27.s, p3/M, z27.s, z28.s\n"
- "fmla z24.s, p3/M, z3.s, z16.s\n"
- "ld1w { z3.s }, p3/Z, [x17, #4, MUL VL]\n"
- "ld1w { z9.s }, p1/Z, [x27, x15, LSL #2]\n"
- "ld1w { z10.s }, p1/Z, [x26, x15, LSL #2]\n"
- "fmax z26.s, p3/M, z26.s, z29.s\n"
- "whilelt p2.s, x14, %x[n_channels]\n"
- "ld1w { z12.s }, p1/Z, [x24, x15, LSL #2]\n"
- "ld1w { z13.s }, p1/Z, [x23, x15, LSL #2]\n"
- "fmla z25.s, p3/M, z5.s, z16.s\n"
- "ld1w { z16.s }, p1/Z, [x20, x15, LSL #2]\n"
- "ld1w { z5.s }, p3/Z, [x17, #6, MUL VL]\n"
- "st1w { z27.s }, p0, [x13, x9, LSL #2]\n"
- "fmla z24.s, p3/M, z7.s, z19.s\n"
- "ld1w { z14.s }, p1/Z, [x22, x15, LSL #2]\n"
- "fmin z26.s, p3/M, z26.s, z28.s\n"
- "fmla z25.s, p3/M, z8.s, z18.s\n"
- "fmla z24.s, p3/M, z6.s, z18.s\n"
- "ld1w { z15.s }, p1/Z, [x21, x15, LSL #2]\n"
- "ld1w { z6.s }, p3/Z, [x17, #7, MUL VL]\n"
- "addvl x17, x17, #16\n"
- "st1w { z26.s }, p0, [x12, x9, LSL #2]\n"
- "fmla z24.s, p3/M, z8.s, z17.s\n"
- "ld1w { z11.s }, p1/Z, [x25, x15, LSL #2]\n"
- "incw x15\n"
- "fmax z25.s, p3/M, z25.s, z29.s\n"
- "ld1w { z7.s }, p3/Z, [x17, #-8, MUL VL]\n"
- "ld1w { z8.s }, p3/Z, [x17, #-7, MUL VL]\n"
- "addvl x17, x17, #-6\n"
- "cmp x15, %x[n_channels]\n"
- "fmin z25.s, p3/M, z25.s, z28.s\n"
- "fmax z24.s, p3/M, z24.s, z29.s\n"
- "fmin z24.s, p3/M, z24.s, z28.s\n"
- "st1w { z25.s }, p0, [x11, x9, LSL #2]\n"
- "st1w { z24.s }, p0, [x10, x9, LSL #2]\n"
+ "ld1w { z20.s }, p3/Z, [x16]\n"
+ "cmp x14, %x[n_channels]\n"
+ "ld1w { z0.s }, p3/Z, [x16, #1, MUL VL]\n"
+ "ld1w { z1.s }, p3/Z, [x16, #2, MUL VL]\n"
+ "fmin z22.s, p3/M, z22.s, z25.s\n"
+ "fmin z21.s, p3/M, z21.s, z25.s\n"
+ "ld1w { z2.s }, p3/Z, [x16, #3, MUL VL]\n"
+ "ld1w { z3.s }, p3/Z, [x16, #4, MUL VL]\n"
+ "st1w { z24.s }, p0, [x13, x28, LSL #2]\n"
+ "ld1w { z4.s }, p3/Z, [x16, #5, MUL VL]\n"
+ "ld1w { z5.s }, p3/Z, [x16, #6, MUL VL]\n"
+ "st1w { z23.s }, p0, [x12, x28, LSL #2]\n"
+ "ld1w { z6.s }, p3/Z, [x16, #7, MUL VL]\n"
+ "addvl x16, x16, #16\n"
+ "st1w { z22.s }, p0, [x11, x28, LSL #2]\n"
+ "ld1w { z7.s }, p3/Z, [x16, #-8, MUL VL]\n"
+ "st1w { z21.s }, p0, [x10, x28, LSL #2]\n"
+ "ld1w { z8.s }, p3/Z, [x16, #-7, MUL VL]\n"
+ "addvl x16, x16, #-6\n"
"blt 1b\n"
"2:" // Channel tail
- "movprfx z27, z30\n fmla z27.s, p3/M, z8.s, z9.s\n"
- "movprfx z26, z30\n fmla z26.s, p3/M, z6.s, z9.s\n"
- "ldr x28, [x16, #0x40]\n"
- "ldr x20, [x16, #0x48]\n"
- "ldr x26, [x16, #0x50]\n"
- "ldr x25, [x16, #0x58]\n"
- "movprfx z25, z30\n fmla z25.s, p3/M, z2.s, z9.s\n"
- "movprfx z24, z30\n fmla z24.s, p3/M, z0.s, z9.s\n"
- "ldr x27, [x16, #0x78]\n"
- "ldr x24, [x16, #0x60]\n"
- "incw x9\n"
+ "movprfx z24, z20\n fmla z24.s, p3/M, z8.s, z9.s\n"
+ "movprfx z23, z20\n fmla z23.s, p3/M, z6.s, z9.s\n"
+ "ldr x21, [x15, #0x40]\n"
+ "ldr x20, [x15, #0x48]\n"
+ "fmla z24.s, p3/M, z0.s, z10.s\n"
+ "fmla z23.s, p3/M, z1.s, z12.s\n"
+ "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x20, [x15, #0x50]\n"
+ "fmla z24.s, p3/M, z1.s, z11.s\n"
+ "fmla z23.s, p3/M, z2.s, z13.s\n"
+ "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z19.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z24.s, p3/M, z3.s, z14.s\n"
+ "fmla z23.s, p3/M, z0.s, z16.s\n"
+ "ldr x20, [x15, #0x58]\n"
+ "ldr x22, [x15, #0x78]\n"
+ "fmla z24.s, p3/M, z4.s, z15.s\n"
+ "fmla z23.s, p3/M, z4.s, z17.s\n"
+ "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x21, [x15, #0x60]\n"
+ "fmla z24.s, p3/M, z2.s, z16.s\n"
+ "fmla z23.s, p3/M, z5.s, z18.s\n"
+ "ldr x20, [x15, #0x80]\n"
+ "ld1w { z18.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "movprfx z22, z20\n fmla z22.s, p3/M, z2.s, z9.s\n"
+ "movprfx z21, z20\n fmla z21.s, p3/M, z0.s, z9.s\n"
+ "ld1w { z20.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x21, [x15, #0x68]\n"
+ "fmla z24.s, p3/M, z5.s, z19.s\n"
+ "fmla z23.s, p3/M, z3.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x22, x9, LSL #2]\n"
+ "ldr x20, [x15, #0x88]\n"
+ "fmla z22.s, p3/M, z3.s, z17.s\n"
+ "fmla z21.s, p3/M, z4.s, z16.s\n"
+ "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z0.s, z18.s\n"
+ "fmla z21.s, p3/M, z1.s, z20.s\n"
+ "ldr x21, [x15, #0x70]\n"
+ "ldr x20, [x15, #0x98]\n"
+ "fmla z22.s, p3/M, z4.s, z17.s\n"
+ "fmla z21.s, p3/M, z5.s, z16.s\n"
+ "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z19.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z24.s, p3/M, z6.s, z18.s\n"
+ "ldr x21, [x15, #0x90]\n"
+ "ldr x20, [x15, #0xa8]\n"
+ "fmla z22.s, p3/M, z1.s, z16.s\n"
+ "fmla z21.s, p3/M, z2.s, z19.s\n"
+ "fmla z24.s, p3/M, z7.s, z16.s\n"
+ "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "ldr x21, [x15, #0xa0]\n"
+ "ldr x20, [x15, #0xb0]\n"
+ "fmla z22.s, p3/M, z6.s, z16.s\n"
+ "fmla z21.s, p3/M, z3.s, z18.s\n"
+ "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
+ "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z7.s, z17.s\n"
+ "fmla z21.s, p3/M, z7.s, z16.s\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "fmla z23.s, p3/M, z7.s, z20.s\n"
+ "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z5.s, z18.s\n"
+ "ldr x20, [x15, #0xc0]\n"
+ "fmla z21.s, p3/M, z6.s, z17.s\n"
+ "fmla z23.s, p3/M, z8.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
+ "fmla z22.s, p3/M, z8.s, z17.s\n"
+ "fmla z21.s, p3/M, z8.s, z16.s\n"
+ "incw x28\n"
"mov p0.b, p2.b\n"
- "ldr x23, [x16, #0x68]\n"
- "ldr x22, [x16, #0x70]\n"
- "fmla z27.s, p3/M, z0.s, z10.s\n"
- "fmla z26.s, p3/M, z1.s, z12.s\n"
- "ld1w { z21.s }, p2/Z, [x20, x14, LSL #2]\n"
- "ldr x21, [x16, #0x88]\n"
- "fmla z27.s, p3/M, z1.s, z11.s\n"
- "ld1w { z18.s }, p2/Z, [x28, x14, LSL #2]\n"
- "ldr x20, [x16, #0x80]\n"
- "fmla z26.s, p3/M, z2.s, z13.s\n"
- "ld1w { z20.s }, p2/Z, [x26, x14, LSL #2]\n"
- "ldr x26, [x16, #0x90]\n"
- "fmla z27.s, p3/M, z3.s, z14.s\n"
- "ld1w { z17.s }, p2/Z, [x25, x14, LSL #2]\n"
- "ldr x25, [x16, #0x98]\n"
- "fmla z26.s, p3/M, z0.s, z16.s\n"
- "fmla z27.s, p3/M, z4.s, z15.s\n"
- "ld1w { z23.s }, p2/Z, [x24, x14, LSL #2]\n"
- "ldr x24, [x16, #0xa0]\n"
- "fmla z25.s, p3/M, z3.s, z17.s\n"
- "ld1w { z22.s }, p2/Z, [x21, x14, LSL #2]\n"
- "fmla z26.s, p3/M, z4.s, z18.s\n"
- "ld1w { z19.s }, p2/Z, [x23, x14, LSL #2]\n"
- "ldr x23, [x16, #0xa8]\n"
- "fmla z27.s, p3/M, z2.s, z16.s\n"
- "ld1w { z18.s }, p2/Z, [x22, x14, LSL #2]\n"
- "ldr x22, [x16, #0xb0]\n"
- "fmla z25.s, p3/M, z0.s, z23.s\n"
- "fmla z26.s, p3/M, z5.s, z21.s\n"
- "ld1w { z17.s }, p2/Z, [x20, x14, LSL #2]\n"
- "ldr x21, [x16, #0xc0]\n"
- "fmla z27.s, p3/M, z5.s, z20.s\n"
- "fmla z26.s, p3/M, z3.s, z20.s\n"
- "ld1w { z16.s }, p2/Z, [x27, x14, LSL #2]\n"
- "ldr x20, [x16, #0xb8]\n"
- "fmla z24.s, p3/M, z4.s, z16.s\n"
- "ld1w { z21.s }, p2/Z, [x24, x14, LSL #2]\n"
- "fmla z25.s, p3/M, z4.s, z19.s\n"
- "ld1w { z20.s }, p2/Z, [x25, x14, LSL #2]\n"
- "fmla z27.s, p3/M, z6.s, z23.s\n"
- "ld1w { z16.s }, p2/Z, [x26, x14, LSL #2]\n"
- "fmla z26.s, p3/M, z7.s, z17.s\n"
- "fmla z24.s, p3/M, z1.s, z17.s\n"
- "fmla z25.s, p3/M, z1.s, z18.s\n"
- "fmla z27.s, p3/M, z7.s, z18.s\n"
- "ld1w { z19.s }, p2/Z, [x23, x14, LSL #2]\n"
- "fmla z26.s, p3/M, z8.s, z20.s\n"
- "fmla z24.s, p3/M, z5.s, z22.s\n"
- "ld1w { z18.s }, p2/Z, [x22, x14, LSL #2]\n"
- "fmla z25.s, p3/M, z6.s, z16.s\n"
- "ld1w { z17.s }, p2/Z, [x20, x14, LSL #2]\n"
- "fmax z27.s, p3/M, z27.s, z29.s\n"
- "fmax z26.s, p3/M, z26.s, z29.s\n"
- "fmla z24.s, p3/M, z2.s, z20.s\n"
- "ld1w { z16.s }, p2/Z, [x21, x14, LSL #2]\n"
- "fmla z25.s, p3/M, z7.s, z21.s\n"
- "fmin z27.s, p3/M, z27.s, z28.s\n"
- "fmin z26.s, p3/M, z26.s, z28.s\n"
- "fmla z24.s, p3/M, z3.s, z19.s\n"
- "st1w { z27.s }, p0, [x13, x9, LSL #2]\n"
- "st1w { z26.s }, p0, [x12, x9, LSL #2]\n"
- "fmla z24.s, p3/M, z7.s, z18.s\n"
- "fmla z25.s, p3/M, z5.s, z19.s\n"
- "fmla z24.s, p3/M, z6.s, z17.s\n"
- "fmla z25.s, p3/M, z8.s, z17.s\n"
- "fmla z24.s, p3/M, z8.s, z16.s\n"
- "fmax z25.s, p3/M, z25.s, z29.s\n"
- "fmin z25.s, p3/M, z25.s, z28.s\n"
- "st1w { z25.s }, p0, [x11, x9, LSL #2]\n"
- "fmax z24.s, p3/M, z24.s, z29.s\n"
- "fmin z24.s, p3/M, z24.s, z28.s\n"
- "st1w { z24.s }, p0, [x10, x9, LSL #2]\n"
+ "fmax z24.s, p3/M, z24.s, z26.s\n"
+ "fmax z23.s, p3/M, z23.s, z26.s\n"
+ "fmax z22.s, p3/M, z22.s, z26.s\n"
+ "fmax z21.s, p3/M, z21.s, z26.s\n"
+ "fmin z24.s, p3/M, z24.s, z25.s\n"
+ "fmin z23.s, p3/M, z23.s, z25.s\n"
+ "st1w { z24.s }, p0, [x13, x28, LSL #2]\n"
+ "fmin z22.s, p3/M, z22.s, z25.s\n"
+ "fmin z21.s, p3/M, z21.s, z25.s\n"
+ "st1w { z23.s }, p0, [x12, x28, LSL #2]\n"
+ "st1w { z22.s }, p0, [x11, x28, LSL #2]\n"
+ "st1w { z21.s }, p0, [x10, x28, LSL #2]\n"
:
: [n_channels] "r" ((unsigned long) n_channels), [offsetof_Args_inptrs] "I" (offsetof(Args, inptrs)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_args_params] "I" (offsetof(Args, params)), [params_struct] "r" (&params_struct)
- : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30"
+ : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26"
);
}