aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp502
1 files changed, 251 insertions, 251 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
index 1770ec182c..7681f346ea 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021, 2023-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -46,8 +46,8 @@ void sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_im
__asm__ __volatile__(
"ptrue p1.b\n"
"mov x9, #0x0\n"
- "ld1rw { z15.s }, p1/Z, [%x[minmax_vals]]\n"
- "ld1rw { z14.s }, p1/Z, [%x[minmax_vals], #4]\n"
+ "ld1rw { z10.s }, p1/Z, [%x[minmax_vals]]\n"
+ "ld1rw { z9.s }, p1/Z, [%x[minmax_vals], #4]\n"
"whilelt p0.s, x9, %x[n_output_channels]\n"
"1:" // Output channel loop
"mov z31.b, #0x0\n"
@@ -55,178 +55,178 @@ void sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_im
"ld1w { z31.s }, p0/Z, [%x[bias], x9, LSL #2]\n"
"2:" // Output channel loop: Load bias: Done
"mov x23, %x[inptrs]\n"
- "ldp x21, x20, [x23], #0x10\n"
"lsr x22, %x[kernel_points], #0x1\n"
"mov z16.d, z31.d\n"
"mov z17.d, z31.d\n"
"mov z18.d, z31.d\n"
- "ld1rqw { z6.s }, p1/Z, [x21]\n"
- "ld1rqw { z5.s }, p1/Z, [x21, #16]\n"
"mov z19.d, z31.d\n"
+ "ld1w { z13.s }, p1/Z, [%x[weights]]\n"
+ "addvl %x[weights], %x[weights], #1\n"
"mov z20.d, z31.d\n"
- "ld1rqw { z1.s }, p1/Z, [x20]\n"
- "ld1rqw { z2.s }, p1/Z, [x20, #16]\n"
"mov z21.d, z31.d\n"
+ "ldp x21, x20, [x23], #0x10\n"
"mov z22.d, z31.d\n"
- "ld1w { z8.s }, p1/Z, [%x[weights]]\n"
- "addvl %x[weights], %x[weights], #1\n"
"mov z23.d, z31.d\n"
"mov z24.d, z31.d\n"
"mov z25.d, z31.d\n"
"mov z26.d, z31.d\n"
"mov z27.d, z31.d\n"
+ "ld1rqw { z7.s }, p1/Z, [x21]\n"
+ "ld1rqw { z6.s }, p1/Z, [x21, #16]\n"
"mov z28.d, z31.d\n"
"mov z29.d, z31.d\n"
+ "ld1rqw { z1.s }, p1/Z, [x20]\n"
+ "ld1rqw { z2.s }, p1/Z, [x20, #16]\n"
"mov z30.d, z31.d\n"
"mov z31.d, z31.d\n"
"cbz x22, 6f\n"
"ldp x21, x20, [x23], #0x10\n"
"subs x22, x22, #0x1\n"
- "ld1rqw { z0.s }, p1/Z, [x21]\n"
- "ld1rqw { z4.s }, p1/Z, [x21, #16]\n"
- "ld1rqw { z7.s }, p1/Z, [x20]\n"
- "ld1rqw { z3.s }, p1/Z, [x20, #16]\n"
- "ld1w { z11.s }, p1/Z, [%x[weights]]\n"
+ "ld1w { z8.s }, p1/Z, [%x[weights]]\n"
"addvl %x[weights], %x[weights], #1\n"
+ "ld1rqw { z5.s }, p1/Z, [x21]\n"
+ "ld1rqw { z3.s }, p1/Z, [x21, #16]\n"
+ "ld1rqw { z0.s }, p1/Z, [x20]\n"
+ "ld1rqw { z4.s }, p1/Z, [x20, #16]\n"
"beq 4f\n"
"3:" // Output channel loop: Kernel loop
"ldp x21, x20, [x23], #0x10\n"
- "fmla z16.s, z8.s, z6.s[0]\n"
- "fmla z17.s, z8.s, z6.s[1]\n"
+ "fmla z16.s, z13.s, z7.s[0]\n"
+ "fmla z17.s, z13.s, z7.s[1]\n"
"subs x22, x22, #0x1\n"
- "fmla z18.s, z8.s, z6.s[2]\n"
- "fmla z19.s, z8.s, z6.s[3]\n"
- "ld1rqw { z6.s }, p1/Z, [x21]\n"
- "fmla z20.s, z8.s, z5.s[0]\n"
- "fmla z21.s, z8.s, z5.s[1]\n"
- "fmla z22.s, z8.s, z5.s[2]\n"
- "fmla z23.s, z8.s, z5.s[3]\n"
- "ld1rqw { z5.s }, p1/Z, [x21, #16]\n"
- "fmla z24.s, z8.s, z1.s[0]\n"
- "fmla z25.s, z8.s, z1.s[1]\n"
- "fmla z26.s, z8.s, z1.s[2]\n"
- "fmla z27.s, z8.s, z1.s[3]\n"
+ "fmla z18.s, z13.s, z7.s[2]\n"
+ "fmla z19.s, z13.s, z7.s[3]\n"
+ "fmla z20.s, z13.s, z6.s[0]\n"
+ "fmla z21.s, z13.s, z6.s[1]\n"
+ "fmla z22.s, z13.s, z6.s[2]\n"
+ "fmla z23.s, z13.s, z6.s[3]\n"
+ "ld1rqw { z7.s }, p1/Z, [x21]\n"
+ "ld1rqw { z6.s }, p1/Z, [x21, #16]\n"
+ "fmla z24.s, z13.s, z1.s[0]\n"
+ "fmla z25.s, z13.s, z1.s[1]\n"
+ "fmla z26.s, z13.s, z1.s[2]\n"
+ "fmla z27.s, z13.s, z1.s[3]\n"
"ld1rqw { z1.s }, p1/Z, [x20]\n"
- "fmla z28.s, z8.s, z2.s[0]\n"
- "fmla z29.s, z8.s, z2.s[1]\n"
- "fmla z30.s, z8.s, z2.s[2]\n"
- "fmla z31.s, z8.s, z2.s[3]\n"
+ "fmla z28.s, z13.s, z2.s[0]\n"
+ "fmla z29.s, z13.s, z2.s[1]\n"
+ "fmla z30.s, z13.s, z2.s[2]\n"
+ "fmla z31.s, z13.s, z2.s[3]\n"
"ld1rqw { z2.s }, p1/Z, [x20, #16]\n"
"ldp x21, x20, [x23], #0x10\n"
- "ld1w { z8.s }, p1/Z, [%x[weights]]\n"
- "fmla z16.s, z11.s, z0.s[0]\n"
- "fmla z17.s, z11.s, z0.s[1]\n"
- "fmla z18.s, z11.s, z0.s[2]\n"
- "fmla z19.s, z11.s, z0.s[3]\n"
- "ld1rqw { z0.s }, p1/Z, [x21]\n"
- "fmla z20.s, z11.s, z4.s[0]\n"
- "fmla z21.s, z11.s, z4.s[1]\n"
- "fmla z22.s, z11.s, z4.s[2]\n"
- "fmla z23.s, z11.s, z4.s[3]\n"
- "ld1rqw { z4.s }, p1/Z, [x21, #16]\n"
- "fmla z24.s, z11.s, z7.s[0]\n"
- "fmla z25.s, z11.s, z7.s[1]\n"
- "fmla z26.s, z11.s, z7.s[2]\n"
- "fmla z27.s, z11.s, z7.s[3]\n"
- "ld1rqw { z7.s }, p1/Z, [x20]\n"
- "fmla z28.s, z11.s, z3.s[0]\n"
- "fmla z29.s, z11.s, z3.s[1]\n"
- "fmla z30.s, z11.s, z3.s[2]\n"
- "fmla z31.s, z11.s, z3.s[3]\n"
- "ld1rqw { z3.s }, p1/Z, [x20, #16]\n"
- "ld1w { z11.s }, p1/Z, [%x[weights], #1, MUL VL]\n"
+ "ld1w { z13.s }, p1/Z, [%x[weights]]\n"
+ "fmla z16.s, z8.s, z5.s[0]\n"
+ "fmla z17.s, z8.s, z5.s[1]\n"
+ "fmla z18.s, z8.s, z5.s[2]\n"
+ "fmla z19.s, z8.s, z5.s[3]\n"
+ "fmla z20.s, z8.s, z3.s[0]\n"
+ "fmla z21.s, z8.s, z3.s[1]\n"
+ "ld1rqw { z5.s }, p1/Z, [x21]\n"
+ "fmla z22.s, z8.s, z3.s[2]\n"
+ "fmla z23.s, z8.s, z3.s[3]\n"
+ "ld1rqw { z3.s }, p1/Z, [x21, #16]\n"
+ "fmla z24.s, z8.s, z0.s[0]\n"
+ "fmla z25.s, z8.s, z0.s[1]\n"
+ "fmla z26.s, z8.s, z0.s[2]\n"
+ "fmla z27.s, z8.s, z0.s[3]\n"
+ "ld1rqw { z0.s }, p1/Z, [x20]\n"
+ "fmla z28.s, z8.s, z4.s[0]\n"
+ "fmla z29.s, z8.s, z4.s[1]\n"
+ "fmla z30.s, z8.s, z4.s[2]\n"
+ "fmla z31.s, z8.s, z4.s[3]\n"
+ "ld1rqw { z4.s }, p1/Z, [x20, #16]\n"
+ "ld1w { z8.s }, p1/Z, [%x[weights], #1, MUL VL]\n"
"addvl %x[weights], %x[weights], #2\n"
"bgt 3b\n"
"4:" // Output channel loop: Kernel loop tail
"tbnz %x[kernel_points], #0, 5f\n"
- "fmla z16.s, z8.s, z6.s[0]\n"
- "fmla z17.s, z8.s, z6.s[1]\n"
+ "fmla z16.s, z13.s, z7.s[0]\n"
+ "fmla z17.s, z13.s, z7.s[1]\n"
"ldr x27, [%x[outptrs], #0x0]\n"
"ldr x26, [%x[outptrs], #0x8]\n"
- "fmla z18.s, z8.s, z6.s[2]\n"
- "fmla z19.s, z8.s, z6.s[3]\n"
+ "fmla z18.s, z13.s, z7.s[2]\n"
+ "fmla z19.s, z13.s, z7.s[3]\n"
"ldr x25, [%x[outptrs], #0x10]\n"
"ldr x24, [%x[outptrs], #0x18]\n"
- "fmla z20.s, z8.s, z5.s[0]\n"
- "fmla z21.s, z8.s, z5.s[1]\n"
+ "fmla z20.s, z13.s, z6.s[0]\n"
+ "fmla z21.s, z13.s, z6.s[1]\n"
"ldr x23, [%x[outptrs], #0x20]\n"
"ldr x22, [%x[outptrs], #0x28]\n"
- "fmla z22.s, z8.s, z5.s[2]\n"
- "fmla z23.s, z8.s, z5.s[3]\n"
+ "fmla z22.s, z13.s, z6.s[2]\n"
+ "fmla z23.s, z13.s, z6.s[3]\n"
"ldr x21, [%x[outptrs], #0x30]\n"
"ldr x20, [%x[outptrs], #0x38]\n"
- "fmla z24.s, z8.s, z1.s[0]\n"
- "fmla z25.s, z8.s, z1.s[1]\n"
- "fmla z26.s, z8.s, z1.s[2]\n"
- "fmla z27.s, z8.s, z1.s[3]\n"
- "fmla z28.s, z8.s, z2.s[0]\n"
- "fmla z29.s, z8.s, z2.s[1]\n"
- "fmla z30.s, z8.s, z2.s[2]\n"
- "fmla z31.s, z8.s, z2.s[3]\n"
- "fmla z16.s, z11.s, z0.s[0]\n"
- "fmla z17.s, z11.s, z0.s[1]\n"
- "fmin z16.s, p1/M, z16.s, z14.s\n"
- "fmin z17.s, p1/M, z17.s, z14.s\n"
- "fmla z18.s, z11.s, z0.s[2]\n"
- "fmla z19.s, z11.s, z0.s[3]\n"
- "fmin z18.s, p1/M, z18.s, z14.s\n"
- "fmin z19.s, p1/M, z19.s, z14.s\n"
- "fmla z20.s, z11.s, z4.s[0]\n"
- "fmla z21.s, z11.s, z4.s[1]\n"
- "fmin z20.s, p1/M, z20.s, z14.s\n"
- "fmin z21.s, p1/M, z21.s, z14.s\n"
- "fmla z22.s, z11.s, z4.s[2]\n"
- "fmla z23.s, z11.s, z4.s[3]\n"
- "fmin z22.s, p1/M, z22.s, z14.s\n"
- "fmin z23.s, p1/M, z23.s, z14.s\n"
- "fmla z24.s, z11.s, z7.s[0]\n"
- "fmla z25.s, z11.s, z7.s[1]\n"
- "fmax z16.s, p1/M, z16.s, z15.s\n"
- "fmax z17.s, p1/M, z17.s, z15.s\n"
- "fmla z26.s, z11.s, z7.s[2]\n"
- "fmla z27.s, z11.s, z7.s[3]\n"
- "fmax z18.s, p1/M, z18.s, z15.s\n"
- "fmax z19.s, p1/M, z19.s, z15.s\n"
- "fmla z28.s, z11.s, z3.s[0]\n"
- "fmla z29.s, z11.s, z3.s[1]\n"
- "fmax z20.s, p1/M, z20.s, z15.s\n"
- "fmax z21.s, p1/M, z21.s, z15.s\n"
- "fmla z30.s, z11.s, z3.s[2]\n"
- "fmla z31.s, z11.s, z3.s[3]\n"
- "fmax z22.s, p1/M, z22.s, z15.s\n"
- "fmax z23.s, p1/M, z23.s, z15.s\n"
- "fmin z24.s, p1/M, z24.s, z14.s\n"
- "fmin z25.s, p1/M, z25.s, z14.s\n"
+ "fmla z24.s, z13.s, z1.s[0]\n"
+ "fmla z25.s, z13.s, z1.s[1]\n"
+ "fmla z26.s, z13.s, z1.s[2]\n"
+ "fmla z27.s, z13.s, z1.s[3]\n"
+ "fmla z28.s, z13.s, z2.s[0]\n"
+ "fmla z29.s, z13.s, z2.s[1]\n"
+ "fmla z30.s, z13.s, z2.s[2]\n"
+ "fmla z31.s, z13.s, z2.s[3]\n"
+ "fmla z16.s, z8.s, z5.s[0]\n"
+ "fmla z17.s, z8.s, z5.s[1]\n"
+ "fmla z18.s, z8.s, z5.s[2]\n"
+ "fmla z19.s, z8.s, z5.s[3]\n"
+ "fmla z20.s, z8.s, z3.s[0]\n"
+ "fmla z21.s, z8.s, z3.s[1]\n"
+ "fmla z22.s, z8.s, z3.s[2]\n"
+ "fmla z23.s, z8.s, z3.s[3]\n"
+ "fmla z24.s, z8.s, z0.s[0]\n"
+ "fmla z25.s, z8.s, z0.s[1]\n"
+ "fmin z16.s, p1/M, z16.s, z9.s\n"
+ "fmin z17.s, p1/M, z17.s, z9.s\n"
+ "fmla z26.s, z8.s, z0.s[2]\n"
+ "fmla z27.s, z8.s, z0.s[3]\n"
+ "fmin z18.s, p1/M, z18.s, z9.s\n"
+ "fmin z19.s, p1/M, z19.s, z9.s\n"
+ "fmla z28.s, z8.s, z4.s[0]\n"
+ "fmla z29.s, z8.s, z4.s[1]\n"
+ "fmin z20.s, p1/M, z20.s, z9.s\n"
+ "fmin z21.s, p1/M, z21.s, z9.s\n"
+ "fmla z30.s, z8.s, z4.s[2]\n"
+ "fmla z31.s, z8.s, z4.s[3]\n"
+ "fmin z22.s, p1/M, z22.s, z9.s\n"
+ "fmin z23.s, p1/M, z23.s, z9.s\n"
+ "fmax z16.s, p1/M, z16.s, z10.s\n"
+ "fmax z17.s, p1/M, z17.s, z10.s\n"
+ "fmax z18.s, p1/M, z18.s, z10.s\n"
+ "fmax z19.s, p1/M, z19.s, z10.s\n"
+ "fmax z20.s, p1/M, z20.s, z10.s\n"
+ "fmax z21.s, p1/M, z21.s, z10.s\n"
+ "fmax z22.s, p1/M, z22.s, z10.s\n"
+ "fmax z23.s, p1/M, z23.s, z10.s\n"
"st1w { z16.s }, p0, [x27, x9, LSL #2]\n"
"ldr x27, [%x[outptrs], #0x40]\n"
- "fmin z26.s, p1/M, z26.s, z14.s\n"
- "fmin z27.s, p1/M, z27.s, z14.s\n"
+ "fmin z24.s, p1/M, z24.s, z9.s\n"
+ "fmin z25.s, p1/M, z25.s, z9.s\n"
"st1w { z17.s }, p0, [x26, x9, LSL #2]\n"
"ldr x26, [%x[outptrs], #0x48]\n"
- "fmin z28.s, p1/M, z28.s, z14.s\n"
- "fmin z29.s, p1/M, z29.s, z14.s\n"
+ "fmin z26.s, p1/M, z26.s, z9.s\n"
+ "fmin z27.s, p1/M, z27.s, z9.s\n"
"st1w { z18.s }, p0, [x25, x9, LSL #2]\n"
"ldr x25, [%x[outptrs], #0x50]\n"
- "fmin z30.s, p1/M, z30.s, z14.s\n"
- "fmin z31.s, p1/M, z31.s, z14.s\n"
+ "fmin z28.s, p1/M, z28.s, z9.s\n"
+ "fmin z29.s, p1/M, z29.s, z9.s\n"
"st1w { z19.s }, p0, [x24, x9, LSL #2]\n"
"ldr x24, [%x[outptrs], #0x58]\n"
+ "fmin z30.s, p1/M, z30.s, z9.s\n"
+ "fmin z31.s, p1/M, z31.s, z9.s\n"
"st1w { z20.s }, p0, [x23, x9, LSL #2]\n"
"ldr x23, [%x[outptrs], #0x60]\n"
- "fmax z24.s, p1/M, z24.s, z15.s\n"
- "fmax z25.s, p1/M, z25.s, z15.s\n"
"st1w { z21.s }, p0, [x22, x9, LSL #2]\n"
"ldr x22, [%x[outptrs], #0x68]\n"
- "fmax z26.s, p1/M, z26.s, z15.s\n"
- "fmax z27.s, p1/M, z27.s, z15.s\n"
+ "fmax z24.s, p1/M, z24.s, z10.s\n"
+ "fmax z25.s, p1/M, z25.s, z10.s\n"
"st1w { z22.s }, p0, [x21, x9, LSL #2]\n"
"ldr x21, [%x[outptrs], #0x70]\n"
- "fmax z28.s, p1/M, z28.s, z15.s\n"
- "fmax z29.s, p1/M, z29.s, z15.s\n"
+ "fmax z26.s, p1/M, z26.s, z10.s\n"
+ "fmax z27.s, p1/M, z27.s, z10.s\n"
"st1w { z23.s }, p0, [x20, x9, LSL #2]\n"
"ldr x20, [%x[outptrs], #0x78]\n"
- "fmax z30.s, p1/M, z30.s, z15.s\n"
- "fmax z31.s, p1/M, z31.s, z15.s\n"
+ "fmax z28.s, p1/M, z28.s, z10.s\n"
+ "fmax z29.s, p1/M, z29.s, z10.s\n"
+ "fmax z30.s, p1/M, z30.s, z10.s\n"
+ "fmax z31.s, p1/M, z31.s, z10.s\n"
"st1w { z24.s }, p0, [x27, x9, LSL #2]\n"
"st1w { z25.s }, p0, [x26, x9, LSL #2]\n"
"st1w { z26.s }, p0, [x25, x9, LSL #2]\n"
@@ -237,117 +237,117 @@ void sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_im
"st1w { z31.s }, p0, [x20, x9, LSL #2]\n"
"b 7f\n"
"5:" // Output channel loop: Odd tail
- "fmla z16.s, z8.s, z6.s[0]\n"
- "fmla z17.s, z8.s, z6.s[1]\n"
+ "fmla z16.s, z13.s, z7.s[0]\n"
+ "fmla z17.s, z13.s, z7.s[1]\n"
"ldp x20, x28, [x23], #0x10\n"
"ldr x27, [%x[outptrs], #0x0]\n"
- "fmla z18.s, z8.s, z6.s[2]\n"
- "fmla z19.s, z8.s, z6.s[3]\n"
- "ld1rqw { z6.s }, p1/Z, [x20]\n"
+ "fmla z18.s, z13.s, z7.s[2]\n"
+ "fmla z19.s, z13.s, z7.s[3]\n"
"ldr x26, [%x[outptrs], #0x8]\n"
- "fmla z20.s, z8.s, z5.s[0]\n"
- "fmla z21.s, z8.s, z5.s[1]\n"
"ldr x25, [%x[outptrs], #0x10]\n"
+ "fmla z20.s, z13.s, z6.s[0]\n"
+ "fmla z21.s, z13.s, z6.s[1]\n"
"ldr x24, [%x[outptrs], #0x18]\n"
- "fmla z22.s, z8.s, z5.s[2]\n"
- "fmla z23.s, z8.s, z5.s[3]\n"
- "ld1rqw { z5.s }, p1/Z, [x20, #16]\n"
"ldr x23, [%x[outptrs], #0x20]\n"
- "fmla z24.s, z8.s, z1.s[0]\n"
- "fmla z25.s, z8.s, z1.s[1]\n"
+ "fmla z22.s, z13.s, z6.s[2]\n"
+ "fmla z23.s, z13.s, z6.s[3]\n"
+ "ld1rqw { z6.s }, p1/Z, [x20]\n"
+ "ld1rqw { z7.s }, p1/Z, [x20, #16]\n"
+ "fmla z24.s, z13.s, z1.s[0]\n"
+ "fmla z25.s, z13.s, z1.s[1]\n"
"ldr x22, [%x[outptrs], #0x28]\n"
"ldr x21, [%x[outptrs], #0x30]\n"
- "fmla z26.s, z8.s, z1.s[2]\n"
- "fmla z27.s, z8.s, z1.s[3]\n"
+ "fmla z26.s, z13.s, z1.s[2]\n"
+ "fmla z27.s, z13.s, z1.s[3]\n"
"ld1rqw { z1.s }, p1/Z, [x28]\n"
"ldr x20, [%x[outptrs], #0x38]\n"
- "fmla z28.s, z8.s, z2.s[0]\n"
- "fmla z29.s, z8.s, z2.s[1]\n"
- "fmla z30.s, z8.s, z2.s[2]\n"
- "fmla z31.s, z8.s, z2.s[3]\n"
- "ld1w { z10.s }, p1/Z, [%x[weights]]\n"
+ "fmla z28.s, z13.s, z2.s[0]\n"
+ "fmla z29.s, z13.s, z2.s[1]\n"
+ "fmla z30.s, z13.s, z2.s[2]\n"
+ "fmla z31.s, z13.s, z2.s[3]\n"
+ "ld1w { z13.s }, p1/Z, [%x[weights]]\n"
"ld1rqw { z2.s }, p1/Z, [x28, #16]\n"
- "fmla z16.s, z11.s, z0.s[0]\n"
- "fmla z17.s, z11.s, z0.s[1]\n"
+ "fmla z16.s, z8.s, z5.s[0]\n"
+ "fmla z17.s, z8.s, z5.s[1]\n"
"addvl %x[weights], %x[weights], #1\n"
- "fmla z18.s, z11.s, z0.s[2]\n"
- "fmla z19.s, z11.s, z0.s[3]\n"
- "fmla z20.s, z11.s, z4.s[0]\n"
- "fmla z21.s, z11.s, z4.s[1]\n"
- "fmla z22.s, z11.s, z4.s[2]\n"
- "fmla z23.s, z11.s, z4.s[3]\n"
- "fmla z24.s, z11.s, z7.s[0]\n"
- "fmla z25.s, z11.s, z7.s[1]\n"
- "fmla z26.s, z11.s, z7.s[2]\n"
- "fmla z27.s, z11.s, z7.s[3]\n"
- "fmla z28.s, z11.s, z3.s[0]\n"
- "fmla z29.s, z11.s, z3.s[1]\n"
- "fmla z30.s, z11.s, z3.s[2]\n"
- "fmla z31.s, z11.s, z3.s[3]\n"
- "fmla z16.s, z10.s, z6.s[0]\n"
- "fmla z17.s, z10.s, z6.s[1]\n"
- "fmin z16.s, p1/M, z16.s, z14.s\n"
- "fmin z17.s, p1/M, z17.s, z14.s\n"
- "fmla z18.s, z10.s, z6.s[2]\n"
- "fmla z19.s, z10.s, z6.s[3]\n"
- "fmin z18.s, p1/M, z18.s, z14.s\n"
- "fmin z19.s, p1/M, z19.s, z14.s\n"
- "fmla z20.s, z10.s, z5.s[0]\n"
- "fmla z21.s, z10.s, z5.s[1]\n"
- "fmin z20.s, p1/M, z20.s, z14.s\n"
- "fmin z21.s, p1/M, z21.s, z14.s\n"
- "fmla z22.s, z10.s, z5.s[2]\n"
- "fmla z23.s, z10.s, z5.s[3]\n"
- "fmin z22.s, p1/M, z22.s, z14.s\n"
- "fmin z23.s, p1/M, z23.s, z14.s\n"
- "fmla z24.s, z10.s, z1.s[0]\n"
- "fmla z25.s, z10.s, z1.s[1]\n"
- "fmax z16.s, p1/M, z16.s, z15.s\n"
- "fmax z17.s, p1/M, z17.s, z15.s\n"
- "fmla z26.s, z10.s, z1.s[2]\n"
- "fmla z27.s, z10.s, z1.s[3]\n"
- "fmax z18.s, p1/M, z18.s, z15.s\n"
- "fmax z19.s, p1/M, z19.s, z15.s\n"
- "fmla z28.s, z10.s, z2.s[0]\n"
- "fmla z29.s, z10.s, z2.s[1]\n"
- "fmax z20.s, p1/M, z20.s, z15.s\n"
- "fmax z21.s, p1/M, z21.s, z15.s\n"
- "fmla z30.s, z10.s, z2.s[2]\n"
- "fmla z31.s, z10.s, z2.s[3]\n"
- "fmax z22.s, p1/M, z22.s, z15.s\n"
- "fmax z23.s, p1/M, z23.s, z15.s\n"
- "fmin z24.s, p1/M, z24.s, z14.s\n"
- "fmin z25.s, p1/M, z25.s, z14.s\n"
+ "fmla z18.s, z8.s, z5.s[2]\n"
+ "fmla z19.s, z8.s, z5.s[3]\n"
+ "fmla z20.s, z8.s, z3.s[0]\n"
+ "fmla z21.s, z8.s, z3.s[1]\n"
+ "fmla z22.s, z8.s, z3.s[2]\n"
+ "fmla z23.s, z8.s, z3.s[3]\n"
+ "fmla z24.s, z8.s, z0.s[0]\n"
+ "fmla z25.s, z8.s, z0.s[1]\n"
+ "fmla z26.s, z8.s, z0.s[2]\n"
+ "fmla z27.s, z8.s, z0.s[3]\n"
+ "fmla z28.s, z8.s, z4.s[0]\n"
+ "fmla z29.s, z8.s, z4.s[1]\n"
+ "fmla z30.s, z8.s, z4.s[2]\n"
+ "fmla z31.s, z8.s, z4.s[3]\n"
+ "fmla z16.s, z13.s, z6.s[0]\n"
+ "fmla z17.s, z13.s, z6.s[1]\n"
+ "fmla z18.s, z13.s, z6.s[2]\n"
+ "fmla z19.s, z13.s, z6.s[3]\n"
+ "fmla z20.s, z13.s, z7.s[0]\n"
+ "fmla z21.s, z13.s, z7.s[1]\n"
+ "fmla z22.s, z13.s, z7.s[2]\n"
+ "fmla z23.s, z13.s, z7.s[3]\n"
+ "fmla z24.s, z13.s, z1.s[0]\n"
+ "fmla z25.s, z13.s, z1.s[1]\n"
+ "fmin z16.s, p1/M, z16.s, z9.s\n"
+ "fmin z17.s, p1/M, z17.s, z9.s\n"
+ "fmla z26.s, z13.s, z1.s[2]\n"
+ "fmla z27.s, z13.s, z1.s[3]\n"
+ "fmin z18.s, p1/M, z18.s, z9.s\n"
+ "fmin z19.s, p1/M, z19.s, z9.s\n"
+ "fmla z28.s, z13.s, z2.s[0]\n"
+ "fmla z29.s, z13.s, z2.s[1]\n"
+ "fmin z20.s, p1/M, z20.s, z9.s\n"
+ "fmin z21.s, p1/M, z21.s, z9.s\n"
+ "fmla z30.s, z13.s, z2.s[2]\n"
+ "fmla z31.s, z13.s, z2.s[3]\n"
+ "fmin z22.s, p1/M, z22.s, z9.s\n"
+ "fmin z23.s, p1/M, z23.s, z9.s\n"
+ "fmax z16.s, p1/M, z16.s, z10.s\n"
+ "fmax z17.s, p1/M, z17.s, z10.s\n"
+ "fmax z18.s, p1/M, z18.s, z10.s\n"
+ "fmax z19.s, p1/M, z19.s, z10.s\n"
+ "fmax z20.s, p1/M, z20.s, z10.s\n"
+ "fmax z21.s, p1/M, z21.s, z10.s\n"
+ "fmax z22.s, p1/M, z22.s, z10.s\n"
+ "fmax z23.s, p1/M, z23.s, z10.s\n"
"st1w { z16.s }, p0, [x27, x9, LSL #2]\n"
"ldr x27, [%x[outptrs], #0x40]\n"
- "fmin z26.s, p1/M, z26.s, z14.s\n"
- "fmin z27.s, p1/M, z27.s, z14.s\n"
+ "fmin z24.s, p1/M, z24.s, z9.s\n"
+ "fmin z25.s, p1/M, z25.s, z9.s\n"
"st1w { z17.s }, p0, [x26, x9, LSL #2]\n"
"ldr x26, [%x[outptrs], #0x48]\n"
- "fmin z28.s, p1/M, z28.s, z14.s\n"
- "fmin z29.s, p1/M, z29.s, z14.s\n"
+ "fmin z26.s, p1/M, z26.s, z9.s\n"
+ "fmin z27.s, p1/M, z27.s, z9.s\n"
"st1w { z18.s }, p0, [x25, x9, LSL #2]\n"
"ldr x25, [%x[outptrs], #0x50]\n"
- "fmin z30.s, p1/M, z30.s, z14.s\n"
- "fmin z31.s, p1/M, z31.s, z14.s\n"
+ "fmin z28.s, p1/M, z28.s, z9.s\n"
+ "fmin z29.s, p1/M, z29.s, z9.s\n"
"st1w { z19.s }, p0, [x24, x9, LSL #2]\n"
"ldr x24, [%x[outptrs], #0x58]\n"
+ "fmin z30.s, p1/M, z30.s, z9.s\n"
+ "fmin z31.s, p1/M, z31.s, z9.s\n"
"st1w { z20.s }, p0, [x23, x9, LSL #2]\n"
"ldr x23, [%x[outptrs], #0x60]\n"
- "fmax z24.s, p1/M, z24.s, z15.s\n"
- "fmax z25.s, p1/M, z25.s, z15.s\n"
"st1w { z21.s }, p0, [x22, x9, LSL #2]\n"
"ldr x22, [%x[outptrs], #0x68]\n"
- "fmax z26.s, p1/M, z26.s, z15.s\n"
- "fmax z27.s, p1/M, z27.s, z15.s\n"
+ "fmax z24.s, p1/M, z24.s, z10.s\n"
+ "fmax z25.s, p1/M, z25.s, z10.s\n"
"st1w { z22.s }, p0, [x21, x9, LSL #2]\n"
"ldr x21, [%x[outptrs], #0x70]\n"
- "fmax z28.s, p1/M, z28.s, z15.s\n"
- "fmax z29.s, p1/M, z29.s, z15.s\n"
+ "fmax z26.s, p1/M, z26.s, z10.s\n"
+ "fmax z27.s, p1/M, z27.s, z10.s\n"
"st1w { z23.s }, p0, [x20, x9, LSL #2]\n"
"ldr x20, [%x[outptrs], #0x78]\n"
- "fmax z30.s, p1/M, z30.s, z15.s\n"
- "fmax z31.s, p1/M, z31.s, z15.s\n"
+ "fmax z28.s, p1/M, z28.s, z10.s\n"
+ "fmax z29.s, p1/M, z29.s, z10.s\n"
+ "fmax z30.s, p1/M, z30.s, z10.s\n"
+ "fmax z31.s, p1/M, z31.s, z10.s\n"
"st1w { z24.s }, p0, [x27, x9, LSL #2]\n"
"st1w { z25.s }, p0, [x26, x9, LSL #2]\n"
"st1w { z26.s }, p0, [x25, x9, LSL #2]\n"
@@ -358,81 +358,81 @@ void sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_im
"st1w { z31.s }, p0, [x20, x9, LSL #2]\n"
"b 7f\n"
"6:" // Output channel loop: Single kernel point
- "fmla z16.s, z8.s, z6.s[0]\n"
- "fmla z17.s, z8.s, z6.s[1]\n"
- "fmin z16.s, p1/M, z16.s, z14.s\n"
- "fmin z17.s, p1/M, z17.s, z14.s\n"
- "fmla z18.s, z8.s, z6.s[2]\n"
- "fmla z19.s, z8.s, z6.s[3]\n"
- "fmin z18.s, p1/M, z18.s, z14.s\n"
- "fmin z19.s, p1/M, z19.s, z14.s\n"
- "fmla z20.s, z8.s, z5.s[0]\n"
- "fmla z21.s, z8.s, z5.s[1]\n"
- "fmin z20.s, p1/M, z20.s, z14.s\n"
- "fmin z21.s, p1/M, z21.s, z14.s\n"
- "fmla z22.s, z8.s, z5.s[2]\n"
- "fmla z23.s, z8.s, z5.s[3]\n"
- "fmin z22.s, p1/M, z22.s, z14.s\n"
- "fmin z23.s, p1/M, z23.s, z14.s\n"
- "fmla z24.s, z8.s, z1.s[0]\n"
- "fmla z25.s, z8.s, z1.s[1]\n"
+ "fmla z16.s, z13.s, z7.s[0]\n"
+ "fmla z17.s, z13.s, z7.s[1]\n"
"ldr x27, [%x[outptrs], #0x0]\n"
"ldr x26, [%x[outptrs], #0x8]\n"
- "fmla z26.s, z8.s, z1.s[2]\n"
- "fmla z27.s, z8.s, z1.s[3]\n"
+ "fmla z18.s, z13.s, z7.s[2]\n"
+ "fmla z19.s, z13.s, z7.s[3]\n"
"ldr x25, [%x[outptrs], #0x10]\n"
"ldr x24, [%x[outptrs], #0x18]\n"
- "fmla z28.s, z8.s, z2.s[0]\n"
- "fmla z29.s, z8.s, z2.s[1]\n"
+ "fmla z20.s, z13.s, z6.s[0]\n"
+ "fmla z21.s, z13.s, z6.s[1]\n"
"ldr x23, [%x[outptrs], #0x20]\n"
"ldr x22, [%x[outptrs], #0x28]\n"
- "fmla z30.s, z8.s, z2.s[2]\n"
- "fmla z31.s, z8.s, z2.s[3]\n"
+ "fmla z22.s, z13.s, z6.s[2]\n"
+ "fmla z23.s, z13.s, z6.s[3]\n"
"ldr x21, [%x[outptrs], #0x30]\n"
"ldr x20, [%x[outptrs], #0x38]\n"
- "fmax z16.s, p1/M, z16.s, z15.s\n"
- "fmax z17.s, p1/M, z17.s, z15.s\n"
+ "fmla z24.s, z13.s, z1.s[0]\n"
+ "fmla z25.s, z13.s, z1.s[1]\n"
+ "fmin z16.s, p1/M, z16.s, z9.s\n"
+ "fmin z17.s, p1/M, z17.s, z9.s\n"
+ "fmla z26.s, z13.s, z1.s[2]\n"
+ "fmla z27.s, z13.s, z1.s[3]\n"
+ "fmin z18.s, p1/M, z18.s, z9.s\n"
+ "fmin z19.s, p1/M, z19.s, z9.s\n"
+ "fmla z28.s, z13.s, z2.s[0]\n"
+ "fmla z29.s, z13.s, z2.s[1]\n"
+ "fmin z20.s, p1/M, z20.s, z9.s\n"
+ "fmin z21.s, p1/M, z21.s, z9.s\n"
+ "fmla z30.s, z13.s, z2.s[2]\n"
+ "fmla z31.s, z13.s, z2.s[3]\n"
+ "fmin z22.s, p1/M, z22.s, z9.s\n"
+ "fmin z23.s, p1/M, z23.s, z9.s\n"
+ "fmax z16.s, p1/M, z16.s, z10.s\n"
+ "fmax z17.s, p1/M, z17.s, z10.s\n"
+ "fmax z18.s, p1/M, z18.s, z10.s\n"
+ "fmax z19.s, p1/M, z19.s, z10.s\n"
+ "fmax z20.s, p1/M, z20.s, z10.s\n"
+ "fmax z21.s, p1/M, z21.s, z10.s\n"
+ "fmax z22.s, p1/M, z22.s, z10.s\n"
+ "fmax z23.s, p1/M, z23.s, z10.s\n"
"st1w { z16.s }, p0, [x27, x9, LSL #2]\n"
"ldr x27, [%x[outptrs], #0x40]\n"
- "fmax z18.s, p1/M, z18.s, z15.s\n"
- "fmax z19.s, p1/M, z19.s, z15.s\n"
+ "fmin z24.s, p1/M, z24.s, z9.s\n"
+ "fmin z25.s, p1/M, z25.s, z9.s\n"
"st1w { z17.s }, p0, [x26, x9, LSL #2]\n"
"ldr x26, [%x[outptrs], #0x48]\n"
- "fmax z20.s, p1/M, z20.s, z15.s\n"
- "fmax z21.s, p1/M, z21.s, z15.s\n"
+ "fmin z26.s, p1/M, z26.s, z9.s\n"
+ "fmin z27.s, p1/M, z27.s, z9.s\n"
"st1w { z18.s }, p0, [x25, x9, LSL #2]\n"
"ldr x25, [%x[outptrs], #0x50]\n"
- "fmax z22.s, p1/M, z22.s, z15.s\n"
- "fmax z23.s, p1/M, z23.s, z15.s\n"
+ "fmin z28.s, p1/M, z28.s, z9.s\n"
+ "fmin z29.s, p1/M, z29.s, z9.s\n"
"st1w { z19.s }, p0, [x24, x9, LSL #2]\n"
"ldr x24, [%x[outptrs], #0x58]\n"
- "fmin z24.s, p1/M, z24.s, z14.s\n"
- "fmin z25.s, p1/M, z25.s, z14.s\n"
+ "fmin z30.s, p1/M, z30.s, z9.s\n"
+ "fmin z31.s, p1/M, z31.s, z9.s\n"
"st1w { z20.s }, p0, [x23, x9, LSL #2]\n"
"ldr x23, [%x[outptrs], #0x60]\n"
- "fmin z26.s, p1/M, z26.s, z14.s\n"
- "fmin z27.s, p1/M, z27.s, z14.s\n"
"st1w { z21.s }, p0, [x22, x9, LSL #2]\n"
"ldr x22, [%x[outptrs], #0x68]\n"
- "fmin z28.s, p1/M, z28.s, z14.s\n"
- "fmin z29.s, p1/M, z29.s, z14.s\n"
+ "fmax z24.s, p1/M, z24.s, z10.s\n"
+ "fmax z25.s, p1/M, z25.s, z10.s\n"
"st1w { z22.s }, p0, [x21, x9, LSL #2]\n"
"ldr x21, [%x[outptrs], #0x70]\n"
- "fmin z30.s, p1/M, z30.s, z14.s\n"
- "fmin z31.s, p1/M, z31.s, z14.s\n"
+ "fmax z26.s, p1/M, z26.s, z10.s\n"
+ "fmax z27.s, p1/M, z27.s, z10.s\n"
"st1w { z23.s }, p0, [x20, x9, LSL #2]\n"
"ldr x20, [%x[outptrs], #0x78]\n"
- "fmax z24.s, p1/M, z24.s, z15.s\n"
- "fmax z25.s, p1/M, z25.s, z15.s\n"
+ "fmax z28.s, p1/M, z28.s, z10.s\n"
+ "fmax z29.s, p1/M, z29.s, z10.s\n"
+ "fmax z30.s, p1/M, z30.s, z10.s\n"
+ "fmax z31.s, p1/M, z31.s, z10.s\n"
"st1w { z24.s }, p0, [x27, x9, LSL #2]\n"
- "fmax z26.s, p1/M, z26.s, z15.s\n"
- "fmax z27.s, p1/M, z27.s, z15.s\n"
"st1w { z25.s }, p0, [x26, x9, LSL #2]\n"
- "fmax z28.s, p1/M, z28.s, z15.s\n"
- "fmax z29.s, p1/M, z29.s, z15.s\n"
"st1w { z26.s }, p0, [x25, x9, LSL #2]\n"
- "fmax z30.s, p1/M, z30.s, z15.s\n"
- "fmax z31.s, p1/M, z31.s, z15.s\n"
"st1w { z27.s }, p0, [x24, x9, LSL #2]\n"
"st1w { z28.s }, p0, [x23, x9, LSL #2]\n"
"st1w { z29.s }, p0, [x22, x9, LSL #2]\n"
@@ -444,7 +444,7 @@ void sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_im
"b.any 1b\n"
: [weights] "+&r" (weights)
: [bias] "r" (bias), [inptrs] "r" (inptrs), [kernel_points] "r" ((uint64_t) kernel_points), [minmax_vals] "r" (minmax_vals), [n_output_channels] "r" ((uint64_t) n_output_channels), [outptrs] "r" (outptrs)
- : "cc", "memory", "p0", "p1", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z10", "z11", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+ : "cc", "memory", "p0", "p1", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z13", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
}