aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_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/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp284
1 files changed, 142 insertions, 142 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
index ebbbd760fc..6b52df7d43 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp
@@ -25,7 +25,7 @@
#include <cstddef>
#include <cstdint>
-#if defined(ARM_COMPUTE_ENABLE_SME2) && defined(__ARM_FP16_ARGS) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+#if defined(ARM_COMPUTE_ENABLE_SME2)
namespace arm_conv {
namespace depthwise {
@@ -96,24 +96,24 @@ void sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
".inst 0x25207810 // ptrue pn8.b\n"
"cnth x13\n"
"whilelt p2.h, XZR, %x[n_channels]\n"
- "ld1rh { z24.h }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
+ "ld1rh { z19.h }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
"ldp x12, x11, [x20, #0x0]\n"
+ "ldp x10, x9, [x20, #0x10]\n"
"cmp x13, %x[n_channels]\n"
- "ld1rh { z27.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
- "sub x10, XZR, x13\n"
- "ldp x9, x28, [x20, #0x10]\n"
- "ld1h { z23.h }, p3/Z, [x14]\n"
+ "ld1rh { z18.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
+ "sub x28, XZR, x13\n"
+ "ld1h { z17.h }, p3/Z, [x14]\n"
"addvl x14, x14, #1\n"
"ldp x27, x26, [x16, #0x0]\n"
+ "ldp x25, x24, [x16, #0x10]\n"
".inst 0xa040a1c0 // ld1h { z0.h-z3.h }, pn8.b/Z, [x14]\n"
"addvl x14, x14, #4\n"
- "ldp x25, x24, [x16, #0x10]\n"
+ "ldp x23, x22, [x16, #0x20]\n"
".inst 0xa040a1c4 // ld1h { z4.h-z7.h }, pn8.b/Z, [x14]\n"
"addvl x14, x14, #4\n"
- "ldp x23, x22, [x16, #0x20]\n"
+ "ldp x21, x20, [x16, #0x30]\n"
"ld1h { z8.h }, p3/Z, [x14]\n"
"addvl x14, x14, #1\n"
- "ldp x21, x20, [x16, #0x30]\n"
"ld1h { z9.h }, p2/Z, [x27, x15, LSL #1]\n"
"ld1h { z10.h }, p2/Z, [x26, x15, LSL #1]\n"
"ld1h { z11.h }, p2/Z, [x25, x15, LSL #1]\n"
@@ -124,187 +124,187 @@ void sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
"ld1h { z16.h }, p2/Z, [x20, x15, LSL #1]\n"
"bge 2f\n"
"1:" // Channel loop
- "movprfx z28, z23\n fmla z28.h, p3/M, z8.h, z9.h\n"
- "movprfx z29, z23\n fmla z29.h, p3/M, z6.h, z9.h\n"
- "ldr x25, [x16, #0x40]\n"
+ "movprfx z28, z17\n fmla z28.h, p3/M, z8.h, z9.h\n"
+ "movprfx z29, z17\n fmla z29.h, p3/M, z6.h, z9.h\n"
+ "ldr x27, [x16, #0x40]\n"
"whilelt p1.h, x13, %x[n_channels]\n"
- "ldr x22, [x16, #0x48]\n"
- "movprfx z30, z23\n fmla z30.h, p3/M, z2.h, z9.h\n"
- "movprfx z31, z23\n fmla z31.h, p3/M, z0.h, z9.h\n"
- "ld1h { z23.h }, p3/Z, [x14]\n"
- "ldr x21, [x16, #0x50]\n"
+ "ldr x26, [x16, #0x48]\n"
+ "movprfx z30, z17\n fmla z30.h, p3/M, z2.h, z9.h\n"
+ "movprfx z31, z17\n fmla z31.h, p3/M, z0.h, z9.h\n"
+ "ld1h { z17.h }, p3/Z, [x14]\n"
+ "ldr x25, [x16, #0x50]\n"
"addvl x14, x14, #1\n"
- "inch x10\n"
- "ldr x20, [x16, #0x58]\n"
+ "inch x28\n"
+ "ldr x24, [x16, #0x58]\n"
"mov p0.b, p2.b\n"
"fmla z28.h, p3/M, z0.h, z10.h\n"
"fmla z29.h, p3/M, z1.h, z12.h\n"
- "ld1h { z20.h }, p2/Z, [x22, x15, LSL #1]\n"
- "ldr x27, [x16, #0x78]\n"
- "ldr x24, [x16, #0x60]\n"
- "ldr x26, [x16, #0x68]\n"
- "ldr x23, [x16, #0x88]\n"
+ "ld1h { z12.h }, p2/Z, [x26, x15, LSL #1]\n"
+ "ldr x20, [x16, #0x78]\n"
+ "ldr x23, [x16, #0x60]\n"
+ "ldr x22, [x16, #0x68]\n"
"fmla z28.h, p3/M, z1.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x27, x15, LSL #1]\n"
"fmla z29.h, p3/M, z2.h, z13.h\n"
- "ld1h { z25.h }, p2/Z, [x25, x15, LSL #1]\n"
- "ldr x22, [x16, #0x80]\n"
- "ld1h { z19.h }, p2/Z, [x21, x15, LSL #1]\n"
+ "ld1h { z13.h }, p2/Z, [x25, x15, LSL #1]\n"
+ "ldr x27, [x16, #0x80]\n"
+ "ldr x26, [x16, #0x88]\n"
"ldr x21, [x16, #0x70]\n"
- "ldr x25, [x16, #0x90]\n"
"fmla z28.h, p3/M, z3.h, z14.h\n"
+ "ld1h { z14.h }, p2/Z, [x24, x15, LSL #1]\n"
"fmla z29.h, p3/M, z0.h, z16.h\n"
- "ld1h { z17.h }, p2/Z, [x20, x15, LSL #1]\n"
- "ldr x20, [x16, #0x98]\n"
- "fmla z30.h, p3/M, z3.h, z17.h\n"
- "ld1h { z22.h }, p2/Z, [x23, x15, LSL #1]\n"
+ "ldr x24, [x16, #0x98]\n"
+ "ldr x25, [x16, #0x90]\n"
+ "fmla z30.h, p3/M, z3.h, z14.h\n"
+ "ld1h { z14.h }, p2/Z, [x26, x15, LSL #1]\n"
"fmla z28.h, p3/M, z4.h, z15.h\n"
- "fmla z29.h, p3/M, z4.h, z25.h\n"
- "ld1h { z18.h }, p2/Z, [x24, x15, LSL #1]\n"
- "ldr x24, [x16, #0xa0]\n"
- "ld1h { z17.h }, p2/Z, [x26, x15, LSL #1]\n"
- "ldr x23, [x16, #0xa8]\n"
- "fmla z30.h, p3/M, z0.h, z18.h\n"
+ "ld1h { z15.h }, p2/Z, [x23, x15, LSL #1]\n"
+ "ldr x23, [x16, #0xa0]\n"
+ "fmla z29.h, p3/M, z4.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x22, x15, LSL #1]\n"
+ "ldr x22, [x16, #0xa8]\n"
"fmla z28.h, p3/M, z2.h, z16.h\n"
- "fmla z29.h, p3/M, z5.h, z20.h\n"
- "ld1h { z25.h }, p2/Z, [x22, x15, LSL #1]\n"
- "ldr x22, [x16, #0xc0]\n"
- "ld1h { z15.h }, p2/Z, [x21, x15, LSL #1]\n"
+ "ld1h { z16.h }, p2/Z, [x21, x15, LSL #1]\n"
"ldr x21, [x16, #0xb0]\n"
- "fmla z30.h, p3/M, z4.h, z17.h\n"
- "ld1h { z20.h }, p2/Z, [x20, x15, LSL #1]\n"
- "fmla z28.h, p3/M, z5.h, z19.h\n"
- "fmla z29.h, p3/M, z3.h, z19.h\n"
- "ld1h { z16.h }, p2/Z, [x27, x15, LSL #1]\n"
+ "fmla z30.h, p3/M, z0.h, z15.h\n"
+ "fmla z29.h, p3/M, z5.h, z12.h\n"
+ "ld1h { z12.h }, p2/Z, [x27, x15, LSL #1]\n"
+ "ldr x27, [x16, #0xc0]\n"
+ "fmla z28.h, p3/M, z5.h, z13.h\n"
+ "fmla z29.h, p3/M, z3.h, z13.h\n"
+ "ld1h { z13.h }, p2/Z, [x20, x15, LSL #1]\n"
"ldr x20, [x16, #0xb8]\n"
- "fmla z31.h, p3/M, z4.h, z16.h\n"
- "fmla z30.h, p3/M, z1.h, z15.h\n"
- "ld1h { z17.h }, p2/Z, [x24, x15, LSL #1]\n"
- "fmla z28.h, p3/M, z6.h, z18.h\n"
- "ld1h { z16.h }, p2/Z, [x25, x15, LSL #1]\n"
- "fmla z29.h, p3/M, z7.h, z25.h\n"
- "fmla z31.h, p3/M, z1.h, z25.h\n"
- "fmla z30.h, p3/M, z6.h, z16.h\n"
- "ld1h { z19.h }, p2/Z, [x20, x15, LSL #1]\n"
- "fmla z28.h, p3/M, z7.h, z15.h\n"
- "ld1h { z16.h }, p2/Z, [x23, x15, LSL #1]\n"
- "fmla z29.h, p3/M, z8.h, z20.h\n"
- "fmla z31.h, p3/M, z5.h, z22.h\n"
- "ld1h { z18.h }, p2/Z, [x21, x15, LSL #1]\n"
- "fmla z30.h, p3/M, z7.h, z17.h\n"
- "fmla z31.h, p3/M, z2.h, z20.h\n"
- "fmla z30.h, p3/M, z5.h, z16.h\n"
- "ld1h { z17.h }, p2/Z, [x22, x15, LSL #1]\n"
+ "fmla z30.h, p3/M, z4.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x24, x15, LSL #1]\n"
+ "fmla z31.h, p3/M, z4.h, z13.h\n"
+ "ld1h { z13.h }, p2/Z, [x23, x15, LSL #1]\n"
+ "fmla z28.h, p3/M, z6.h, z15.h\n"
+ "ld1h { z15.h }, p2/Z, [x25, x15, LSL #1]\n"
+ "fmla z29.h, p3/M, z7.h, z12.h\n"
+ "fmla z30.h, p3/M, z1.h, z16.h\n"
+ "fmla z31.h, p3/M, z1.h, z12.h\n"
+ "fmla z28.h, p3/M, z7.h, z16.h\n"
+ "ld1h { z16.h }, p2/Z, [x22, x15, LSL #1]\n"
+ "fmla z30.h, p3/M, z6.h, z15.h\n"
+ "ld1h { z15.h }, p2/Z, [x20, x15, LSL #1]\n"
+ "fmla z29.h, p3/M, z8.h, z11.h\n"
+ "fmla z31.h, p3/M, z5.h, z14.h\n"
+ "ld1h { z14.h }, p2/Z, [x21, x15, LSL #1]\n"
+ "fmla z30.h, p3/M, z7.h, z13.h\n"
+ "fmla z31.h, p3/M, z2.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x27, x15, LSL #1]\n"
"ldp x27, x26, [x16, #0x0]\n"
- "ldp x25, x24, [x16, #0x10]\n"
"inch x15\n"
- "ldp x23, x22, [x16, #0x20]\n"
+ "ldp x25, x24, [x16, #0x10]\n"
"whilelt p2.h, x15, %x[n_channels]\n"
+ "ldp x23, x22, [x16, #0x20]\n"
+ "fmla z30.h, p3/M, z5.h, z16.h\n"
"ldp x21, x20, [x16, #0x30]\n"
"ld1h { z9.h }, p1/Z, [x27, x13, LSL #1]\n"
"fmla z31.h, p3/M, z3.h, z16.h\n"
- "fmla z30.h, p3/M, z8.h, z19.h\n"
"ld1h { z10.h }, p1/Z, [x26, x13, LSL #1]\n"
"ld1h { z12.h }, p1/Z, [x24, x13, LSL #1]\n"
+ "fmla z30.h, p3/M, z8.h, z15.h\n"
"ld1h { z13.h }, p1/Z, [x23, x13, LSL #1]\n"
- "ld1h { z16.h }, p1/Z, [x20, x13, LSL #1]\n"
- "fmla z31.h, p3/M, z7.h, z18.h\n"
+ "fmla z31.h, p3/M, z7.h, z14.h\n"
"ld1h { z14.h }, p1/Z, [x22, x13, LSL #1]\n"
+ "ld1h { z16.h }, p1/Z, [x20, x13, LSL #1]\n"
".inst 0xa040a1c0 // ld1h { z0.h-z3.h }, pn8.b/Z, [x14]\n"
"addvl x14, x14, #4\n"
- "fmla z31.h, p3/M, z6.h, z19.h\n"
+ "fmla z31.h, p3/M, z6.h, z15.h\n"
"ld1h { z15.h }, p1/Z, [x21, x13, LSL #1]\n"
".inst 0xa040a1c4 // ld1h { z4.h-z7.h }, pn8.b/Z, [x14]\n"
"addvl x14, x14, #4\n"
- "fmla z31.h, p3/M, z8.h, z17.h\n"
+ "fmla z31.h, p3/M, z8.h, z11.h\n"
"ld1h { z11.h }, p1/Z, [x25, x13, LSL #1]\n"
"inch x13\n"
"cmp x13, %x[n_channels]\n"
"ld1h { z8.h }, p3/Z, [x14]\n"
"addvl x14, x14, #1\n"
- ".inst 0xc17bcb1c // fclamp { z28.h-z31.h }, z24.h, z27.h\n"
- "st1h { z28.h }, p0, [x12, x10, LSL #1]\n"
- "st1h { z29.h }, p0, [x11, x10, LSL #1]\n"
- "st1h { z30.h }, p0, [x9, x10, LSL #1]\n"
- "st1h { z31.h }, p0, [x28, x10, LSL #1]\n"
+ ".inst 0xc172ca7c // fclamp { z28.h-z31.h }, z19.h, z18.h\n"
+ "st1h { z28.h }, p0, [x12, x28, LSL #1]\n"
+ "st1h { z29.h }, p0, [x11, x28, LSL #1]\n"
+ "st1h { z30.h }, p0, [x10, x28, LSL #1]\n"
+ "st1h { z31.h }, p0, [x9, x28, LSL #1]\n"
"blt 1b\n"
"2:" // Channel tail
- "movprfx z28, z23\n fmla z28.h, p3/M, z8.h, z9.h\n"
- "movprfx z29, z23\n fmla z29.h, p3/M, z6.h, z9.h\n"
- "ldr x25, [x16, #0x40]\n"
- "inch x10\n"
- "ldr x22, [x16, #0x48]\n"
- "movprfx z30, z23\n fmla z30.h, p3/M, z2.h, z9.h\n"
- "movprfx z31, z23\n fmla z31.h, p3/M, z0.h, z9.h\n"
+ "movprfx z28, z17\n fmla z28.h, p3/M, z8.h, z9.h\n"
+ "movprfx z29, z17\n fmla z29.h, p3/M, z6.h, z9.h\n"
+ "ldr x27, [x16, #0x40]\n"
+ "inch x28\n"
+ "ldr x26, [x16, #0x48]\n"
+ "movprfx z30, z17\n fmla z30.h, p3/M, z2.h, z9.h\n"
+ "movprfx z31, z17\n fmla z31.h, p3/M, z0.h, z9.h\n"
"mov p0.b, p2.b\n"
- "ldr x21, [x16, #0x50]\n"
- "ldr x20, [x16, #0x58]\n"
+ "ldr x25, [x16, #0x50]\n"
+ "ldr x24, [x16, #0x58]\n"
"fmla z28.h, p3/M, z0.h, z10.h\n"
"fmla z29.h, p3/M, z1.h, z12.h\n"
- "ld1h { z21.h }, p2/Z, [x22, x15, LSL #1]\n"
- "ldr x27, [x16, #0x78]\n"
- "ldr x24, [x16, #0x60]\n"
- "ldr x26, [x16, #0x68]\n"
- "ldr x23, [x16, #0x88]\n"
+ "ld1h { z12.h }, p2/Z, [x26, x15, LSL #1]\n"
+ "ldr x20, [x16, #0x78]\n"
+ "ldr x23, [x16, #0x60]\n"
+ "ldr x22, [x16, #0x68]\n"
"fmla z28.h, p3/M, z1.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x27, x15, LSL #1]\n"
"fmla z29.h, p3/M, z2.h, z13.h\n"
- "ld1h { z18.h }, p2/Z, [x25, x15, LSL #1]\n"
- "ldr x22, [x16, #0x80]\n"
- "ld1h { z20.h }, p2/Z, [x21, x15, LSL #1]\n"
+ "ld1h { z13.h }, p2/Z, [x25, x15, LSL #1]\n"
+ "ldr x27, [x16, #0x80]\n"
+ "ldr x26, [x16, #0x88]\n"
"ldr x21, [x16, #0x70]\n"
- "ldr x25, [x16, #0x90]\n"
"fmla z28.h, p3/M, z3.h, z14.h\n"
+ "ld1h { z14.h }, p2/Z, [x24, x15, LSL #1]\n"
"fmla z29.h, p3/M, z0.h, z16.h\n"
- "ld1h { z17.h }, p2/Z, [x20, x15, LSL #1]\n"
- "ldr x20, [x16, #0x98]\n"
- "fmla z30.h, p3/M, z3.h, z17.h\n"
- "ld1h { z23.h }, p2/Z, [x23, x15, LSL #1]\n"
+ "ldr x24, [x16, #0x98]\n"
+ "ldr x25, [x16, #0x90]\n"
+ "fmla z30.h, p3/M, z3.h, z14.h\n"
+ "ld1h { z14.h }, p2/Z, [x26, x15, LSL #1]\n"
"fmla z28.h, p3/M, z4.h, z15.h\n"
- "fmla z29.h, p3/M, z4.h, z18.h\n"
- "ld1h { z22.h }, p2/Z, [x24, x15, LSL #1]\n"
- "ldr x24, [x16, #0xa0]\n"
- "ld1h { z19.h }, p2/Z, [x26, x15, LSL #1]\n"
- "ldr x23, [x16, #0xa8]\n"
- "fmla z30.h, p3/M, z0.h, z22.h\n"
+ "ld1h { z15.h }, p2/Z, [x23, x15, LSL #1]\n"
+ "ldr x23, [x16, #0xa0]\n"
+ "fmla z29.h, p3/M, z4.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x22, x15, LSL #1]\n"
+ "ldr x22, [x16, #0xa8]\n"
"fmla z28.h, p3/M, z2.h, z16.h\n"
- "fmla z29.h, p3/M, z5.h, z21.h\n"
- "ld1h { z18.h }, p2/Z, [x22, x15, LSL #1]\n"
- "ldr x22, [x16, #0xc0]\n"
- "ld1h { z17.h }, p2/Z, [x21, x15, LSL #1]\n"
+ "ld1h { z16.h }, p2/Z, [x21, x15, LSL #1]\n"
"ldr x21, [x16, #0xb0]\n"
- "fmla z30.h, p3/M, z4.h, z19.h\n"
- "ld1h { z21.h }, p2/Z, [x20, x15, LSL #1]\n"
- "fmla z28.h, p3/M, z5.h, z20.h\n"
- "fmla z29.h, p3/M, z3.h, z20.h\n"
- "ld1h { z16.h }, p2/Z, [x27, x15, LSL #1]\n"
+ "fmla z30.h, p3/M, z0.h, z15.h\n"
+ "fmla z29.h, p3/M, z5.h, z12.h\n"
+ "ld1h { z12.h }, p2/Z, [x27, x15, LSL #1]\n"
+ "ldr x27, [x16, #0xc0]\n"
+ "fmla z28.h, p3/M, z5.h, z13.h\n"
+ "fmla z29.h, p3/M, z3.h, z13.h\n"
+ "ld1h { z13.h }, p2/Z, [x20, x15, LSL #1]\n"
"ldr x20, [x16, #0xb8]\n"
- "fmla z31.h, p3/M, z4.h, z16.h\n"
- "fmla z30.h, p3/M, z1.h, z17.h\n"
- "ld1h { z20.h }, p2/Z, [x24, x15, LSL #1]\n"
- "fmla z28.h, p3/M, z6.h, z22.h\n"
- "ld1h { z16.h }, p2/Z, [x25, x15, LSL #1]\n"
- "fmla z29.h, p3/M, z7.h, z18.h\n"
- "fmla z31.h, p3/M, z1.h, z18.h\n"
- "fmla z30.h, p3/M, z6.h, z16.h\n"
- "ld1h { z19.h }, p2/Z, [x20, x15, LSL #1]\n"
- "fmla z28.h, p3/M, z7.h, z17.h\n"
- "ld1h { z18.h }, p2/Z, [x23, x15, LSL #1]\n"
- "fmla z29.h, p3/M, z8.h, z21.h\n"
- "fmla z31.h, p3/M, z5.h, z23.h\n"
- "ld1h { z17.h }, p2/Z, [x21, x15, LSL #1]\n"
- "fmla z30.h, p3/M, z7.h, z20.h\n"
- "fmla z31.h, p3/M, z2.h, z21.h\n"
- "fmla z30.h, p3/M, z5.h, z18.h\n"
+ "fmla z30.h, p3/M, z4.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x24, x15, LSL #1]\n"
+ "fmla z31.h, p3/M, z4.h, z13.h\n"
+ "ld1h { z13.h }, p2/Z, [x23, x15, LSL #1]\n"
+ "fmla z28.h, p3/M, z6.h, z15.h\n"
+ "ld1h { z15.h }, p2/Z, [x25, x15, LSL #1]\n"
+ "fmla z29.h, p3/M, z7.h, z12.h\n"
+ "fmla z30.h, p3/M, z1.h, z16.h\n"
+ "fmla z31.h, p3/M, z1.h, z12.h\n"
+ "fmla z28.h, p3/M, z7.h, z16.h\n"
"ld1h { z16.h }, p2/Z, [x22, x15, LSL #1]\n"
- "fmla z31.h, p3/M, z3.h, z18.h\n"
- "fmla z30.h, p3/M, z8.h, z19.h\n"
- "fmla z31.h, p3/M, z7.h, z17.h\n"
- "fmla z31.h, p3/M, z6.h, z19.h\n"
- "fmla z31.h, p3/M, z8.h, z16.h\n"
- ".inst 0xc17bcb1c // fclamp { z28.h-z31.h }, z24.h, z27.h\n"
- "st1h { z28.h }, p0, [x12, x10, LSL #1]\n"
- "st1h { z29.h }, p0, [x11, x10, LSL #1]\n"
- "st1h { z30.h }, p0, [x9, x10, LSL #1]\n"
- "st1h { z31.h }, p0, [x28, x10, LSL #1]\n"
+ "fmla z30.h, p3/M, z6.h, z15.h\n"
+ "ld1h { z15.h }, p2/Z, [x20, x15, LSL #1]\n"
+ "fmla z29.h, p3/M, z8.h, z11.h\n"
+ "fmla z31.h, p3/M, z5.h, z14.h\n"
+ "ld1h { z14.h }, p2/Z, [x21, x15, LSL #1]\n"
+ "fmla z30.h, p3/M, z7.h, z13.h\n"
+ "fmla z31.h, p3/M, z2.h, z11.h\n"
+ "ld1h { z11.h }, p2/Z, [x27, x15, LSL #1]\n"
+ "fmla z30.h, p3/M, z5.h, z16.h\n"
+ "fmla z31.h, p3/M, z3.h, z16.h\n"
+ "fmla z30.h, p3/M, z8.h, z15.h\n"
+ "fmla z31.h, p3/M, z7.h, z14.h\n"
+ "fmla z31.h, p3/M, z6.h, z15.h\n"
+ "fmla z31.h, p3/M, z8.h, z11.h\n"
+ ".inst 0xc172ca7c // fclamp { z28.h-z31.h }, z19.h, z18.h\n"
+ "st1h { z28.h }, p0, [x12, x28, LSL #1]\n"
+ "st1h { z29.h }, p0, [x11, x28, LSL #1]\n"
+ "st1h { z30.h }, p0, [x10, x28, LSL #1]\n"
+ "st1h { z31.h }, p0, [x9, x28, LSL #1]\n"
".inst 0xd503467f // SMSTOP\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)
@@ -315,4 +315,4 @@ void sme2_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
} // namespace depthwise
} // namespace arm_conv
-#endif // defined(ARM_COMPUTE_ENABLE_SME2) && defined(__ARM_FP16_ARGS) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+#endif // defined(ARM_COMPUTE_ENABLE_SME2)