aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp466
1 files changed, 233 insertions, 233 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
index ce0ae29756..d15a3a8377 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
@@ -22,11 +22,11 @@
* SOFTWARE.
*/
-#if defined(ARM_COMPUTE_ENABLE_SME2)
-
#include <cstddef>
#include <cstdint>
+#if defined(ARM_COMPUTE_ENABLE_SME2)
+
namespace arm_conv {
namespace depthwise {
@@ -170,11 +170,11 @@ void sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"ldr x22, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
"mul x21, x2, x22\n" // offset = tile_i * ld_output_row
"mov x20, #0x3\n"
- "ld1w { z18.s }, p3/Z, [x17]\n"
+ "ld1w { z24.s }, p3/Z, [x17]\n"
"ldr x27, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
"madd x21, x3, x27, x21\n" // offset += tile_j * ld_output_col
"mul x21, x21, x20\n" // offset *= output_tile_size
- "ld1rw { z17.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
+ "ld1rw { z26.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
"ldr x26, [%x[params_struct], %[offsetof_args_outptr]]\n"
"addvl x17, x17, #1\n"
"add x26, x26, x21, LSL #2\n" // outptrs[0] += offset * sizeof(float)
@@ -184,7 +184,7 @@ void sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
".inst 0xa040c224 // ld1w { z4.s-z7.s }, pn8.b/Z, [x17]\n"
"add x24, x26, x22, LSL #2\n"
"whilelt p2.s, XZR, %x[n_channels]\n"
- "ld1rw { z16.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
+ "ld1rw { z14.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
"addvl x17, x17, #4\n"
"cmp x25, %x[n_channels]\n"
"ld1w { z8.s }, p3/Z, [x17]\n"
@@ -200,275 +200,275 @@ void sme2_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"ld1w { z13.s }, p2/Z, [x6, x8, LSL #2]\n"
"bge 4f\n"
"3:" // Tile loop: Channel loop
- "movprfx z24, z18\n fmla z24.s, p3/M, z7.s, z9.s\n"
- "movprfx z23, z18\n fmla z23.s, p3/M, z8.s, z9.s\n"
+ "movprfx z28, z24\n fmla z28.s, p3/M, z7.s, z9.s\n"
+ "movprfx z27, z24\n fmla z27.s, p3/M, z8.s, z9.s\n"
"whilelt p1.s, x25, %x[n_channels]\n"
"incw x21\n"
- "movprfx z25, z18\n fmla z25.s, p3/M, z6.s, z9.s\n"
- "fmla z24.s, p3/M, z4.s, z13.s\n"
+ "movprfx z29, z24\n fmla z29.s, p3/M, z6.s, z9.s\n"
+ "fmla z28.s, p3/M, z4.s, z13.s\n"
"incw x25\n"
"mov p0.b, p2.b\n"
- "movprfx z26, z18\n fmla z26.s, p3/M, z5.s, z9.s\n"
- "movprfx z27, z18\n fmla z27.s, p3/M, z4.s, z9.s\n"
+ "movprfx z30, z24\n fmla z30.s, p3/M, z5.s, z9.s\n"
+ "movprfx z31, z24\n fmla z31.s, p3/M, z4.s, z9.s\n"
"incw x20\n"
- "movprfx z28, z18\n fmla z28.s, p3/M, z3.s, z9.s\n"
- "fmla z23.s, p3/M, z0.s, z10.s\n"
+ "movprfx z20, z24\n fmla z20.s, p3/M, z3.s, z9.s\n"
+ "fmla z27.s, p3/M, z0.s, z10.s\n"
"ld1w { z10.s }, p2/Z, [x7, x15, LSL #2]\n"
- "fmla z25.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x7, x4, LSL #2]\n"
- "movprfx z29, z18\n fmla z29.s, p3/M, z2.s, z9.s\n"
- "fmla z24.s, p3/M, z6.s, z11.s\n"
- "movprfx z31, z18\n fmla z31.s, p3/M, z0.s, z9.s\n"
- "fmla z23.s, p3/M, z5.s, z13.s\n"
- "fmla z25.s, p3/M, z3.s, z13.s\n"
- "fmla z26.s, p3/M, z2.s, z13.s\n"
- "fmla z27.s, p3/M, z1.s, z13.s\n"
- "fmla z28.s, p3/M, z0.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x5, x4, LSL #2]\n"
- "fmla z29.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x13, LSL #2]\n"
- "movprfx z30, z18\n fmla z30.s, p3/M, z1.s, z9.s\n"
- "fmla z24.s, p3/M, z0.s, z13.s\n"
- "fmla z31.s, p3/M, z8.s, z12.s\n"
- "fmla z23.s, p3/M, z7.s, z11.s\n"
- "ld1w { z12.s }, p2/Z, [x5, x15, LSL #2]\n"
- "fmla z30.s, p3/M, z0.s, z11.s\n"
- "fmla z26.s, p3/M, z4.s, z11.s\n"
- "ld1w { z18.s }, p3/Z, [x17]\n"
+ "fmla z29.s, p3/M, z2.s, z11.s\n"
+ "ld1w { z19.s }, p2/Z, [x7, x4, LSL #2]\n"
+ "movprfx z21, z24\n fmla z21.s, p3/M, z2.s, z9.s\n"
+ "fmla z28.s, p3/M, z6.s, z19.s\n"
+ "movprfx z23, z24\n fmla z23.s, p3/M, z0.s, z9.s\n"
+ "fmla z27.s, p3/M, z5.s, z13.s\n"
+ "fmla z29.s, p3/M, z3.s, z13.s\n"
+ "fmla z30.s, p3/M, z2.s, z13.s\n"
+ "fmla z31.s, p3/M, z1.s, z13.s\n"
+ "fmla z20.s, p3/M, z0.s, z13.s\n"
+ "ld1w { z18.s }, p2/Z, [x5, x4, LSL #2]\n"
+ "fmla z21.s, p3/M, z6.s, z12.s\n"
+ "ld1w { z15.s }, p2/Z, [x14, x13, LSL #2]\n"
+ "movprfx z22, z24\n fmla z22.s, p3/M, z1.s, z9.s\n"
+ "fmla z28.s, p3/M, z0.s, z18.s\n"
+ "fmla z23.s, p3/M, z8.s, z15.s\n"
+ "fmla z27.s, p3/M, z7.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x5, x15, LSL #2]\n"
+ "fmla z22.s, p3/M, z0.s, z19.s\n"
+ "fmla z30.s, p3/M, z4.s, z19.s\n"
+ "ld1w { z24.s }, p3/Z, [x17]\n"
"addvl x17, x17, #1\n"
- "fmla z27.s, p3/M, z3.s, z11.s\n"
- "fmla z29.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x6]\n"
- "fmla z24.s, p3/M, z2.s, z12.s\n"
- "fmla z25.s, p3/M, z1.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x16]\n"
- "fmla z28.s, p3/M, z4.s, z10.s\n"
- "fmla z23.s, p3/M, z1.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x6, x13, LSL #2]\n"
- "fmla z30.s, p3/M, z2.s, z10.s\n"
- "fmla z31.s, p3/M, z1.s, z10.s\n"
- "fmla z24.s, p3/M, z8.s, z10.s\n"
- "fmla z25.s, p3/M, z7.s, z10.s\n"
- "fmla z27.s, p3/M, z5.s, z10.s\n"
- "fmla z26.s, p3/M, z0.s, z11.s\n"
- "ld1w { z10.s }, p2/Z, [x16, x8, LSL #2]\n"
- "fmla z28.s, p3/M, z2.s, z13.s\n"
- "fmla z29.s, p3/M, z3.s, z12.s\n"
- "fmla z30.s, p3/M, z4.s, z10.s\n"
- "fmla z31.s, p3/M, z3.s, z10.s\n"
- "fmla z23.s, p3/M, z3.s, z11.s\n"
- "fmla z25.s, p3/M, z5.s, z13.s\n"
- "ld1w { z11.s }, p2/Z, [x16, x13, LSL #2]\n"
- "fmla z26.s, p3/M, z6.s, z12.s\n"
- "fmla z27.s, p3/M, z7.s, z10.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x4, LSL #2]\n"
- "fmla z28.s, p3/M, z6.s, z10.s\n"
- "fmla z29.s, p3/M, z5.s, z10.s\n"
- "ld1w { z12.s }, p2/Z, [x6, x4, LSL #2]\n"
- "fmla z31.s, p3/M, z5.s, z11.s\n"
- "fmla z30.s, p3/M, z6.s, z13.s\n"
- "fmla z26.s, p3/M, z8.s, z10.s\n"
- "fmla z28.s, p3/M, z8.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x6, x15, LSL #2]\n"
+ "fmla z31.s, p3/M, z3.s, z19.s\n"
+ "fmla z21.s, p3/M, z1.s, z19.s\n"
+ "ld1w { z17.s }, p2/Z, [x6]\n"
+ "fmla z28.s, p3/M, z2.s, z16.s\n"
+ "fmla z29.s, p3/M, z1.s, z16.s\n"
+ "ld1w { z16.s }, p2/Z, [x16]\n"
+ "fmla z20.s, p3/M, z4.s, z10.s\n"
+ "fmla z27.s, p3/M, z1.s, z18.s\n"
+ "ld1w { z9.s }, p2/Z, [x6, x13, LSL #2]\n"
+ "fmla z22.s, p3/M, z2.s, z10.s\n"
+ "fmla z23.s, p3/M, z1.s, z10.s\n"
+ "fmla z28.s, p3/M, z8.s, z10.s\n"
+ "fmla z29.s, p3/M, z7.s, z10.s\n"
+ "fmla z31.s, p3/M, z5.s, z10.s\n"
+ "fmla z30.s, p3/M, z0.s, z17.s\n"
+ "ld1w { z19.s }, p2/Z, [x16, x8, LSL #2]\n"
+ "fmla z20.s, p3/M, z2.s, z9.s\n"
+ "fmla z21.s, p3/M, z3.s, z16.s\n"
+ "fmla z22.s, p3/M, z4.s, z19.s\n"
+ "fmla z23.s, p3/M, z3.s, z19.s\n"
+ "fmla z27.s, p3/M, z3.s, z17.s\n"
+ "fmla z29.s, p3/M, z5.s, z9.s\n"
+ "ld1w { z17.s }, p2/Z, [x16, x13, LSL #2]\n"
+ "fmla z30.s, p3/M, z6.s, z16.s\n"
+ "fmla z31.s, p3/M, z7.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x14, x4, LSL #2]\n"
+ "fmla z20.s, p3/M, z6.s, z19.s\n"
+ "fmla z21.s, p3/M, z5.s, z19.s\n"
+ "ld1w { z18.s }, p2/Z, [x6, x4, LSL #2]\n"
+ "fmla z23.s, p3/M, z5.s, z17.s\n"
+ "fmla z22.s, p3/M, z6.s, z16.s\n"
+ "fmla z30.s, p3/M, z8.s, z19.s\n"
+ "fmla z20.s, p3/M, z8.s, z17.s\n"
+ "ld1w { z17.s }, p2/Z, [x6, x15, LSL #2]\n"
"addvl x6, x6, #1\n"
- "fmla z29.s, p3/M, z7.s, z13.s\n"
- "fmla z24.s, p3/M, z3.s, z12.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x15, LSL #2]\n"
- "fmla z27.s, p3/M, z0.s, z12.s\n"
- "fmla z23.s, p3/M, z4.s, z12.s\n"
- "fmla z30.s, p3/M, z8.s, z13.s\n"
- "fmla z31.s, p3/M, z7.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x16, x15, LSL #2]\n"
- "fmla z26.s, p3/M, z1.s, z12.s\n"
- "fmla z24.s, p3/M, z5.s, z11.s\n"
- "ld1w { z12.s }, p2/Z, [x16, x4, LSL #2]\n"
+ "fmla z21.s, p3/M, z7.s, z16.s\n"
+ "fmla z28.s, p3/M, z3.s, z18.s\n"
+ "ld1w { z16.s }, p2/Z, [x14, x15, LSL #2]\n"
+ "fmla z31.s, p3/M, z0.s, z18.s\n"
+ "fmla z27.s, p3/M, z4.s, z18.s\n"
+ "fmla z22.s, p3/M, z8.s, z16.s\n"
+ "fmla z23.s, p3/M, z7.s, z16.s\n"
+ "ld1w { z19.s }, p2/Z, [x16, x15, LSL #2]\n"
+ "fmla z30.s, p3/M, z1.s, z18.s\n"
+ "fmla z28.s, p3/M, z5.s, z17.s\n"
+ "ld1w { z11.s }, p2/Z, [x16, x4, LSL #2]\n"
"addvl x16, x16, #1\n"
- "fmla z25.s, p3/M, z4.s, z11.s\n"
- "fmla z27.s, p3/M, z2.s, z11.s\n"
- "fmla z28.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x5, x8, LSL #2]\n"
- "fmla z29.s, p3/M, z4.s, z12.s\n"
+ "fmla z29.s, p3/M, z4.s, z17.s\n"
+ "fmla z31.s, p3/M, z2.s, z17.s\n"
+ "fmla z20.s, p3/M, z1.s, z17.s\n"
+ "ld1w { z16.s }, p2/Z, [x5, x8, LSL #2]\n"
+ "fmla z21.s, p3/M, z4.s, z11.s\n"
"addvl x5, x5, #1\n"
- "fmla z30.s, p3/M, z3.s, z12.s\n"
- "fmla z23.s, p3/M, z2.s, z11.s\n"
+ "fmla z22.s, p3/M, z3.s, z11.s\n"
+ "fmla z27.s, p3/M, z2.s, z16.s\n"
"ld1w { z10.s }, p1/Z, [x5]\n"
- "fmla z31.s, p3/M, z4.s, z13.s\n"
- "fmla z26.s, p3/M, z7.s, z12.s\n"
- "fmla z27.s, p3/M, z6.s, z12.s\n"
- "fmla z24.s, p3/M, z1.s, z11.s\n"
- "ld1w { z12.s }, p2/Z, [x7]\n"
- "fmla z25.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x7, x13, LSL #2]\n"
- "fmla z28.s, p3/M, z7.s, z13.s\n"
+ "fmla z23.s, p3/M, z4.s, z19.s\n"
+ "fmla z30.s, p3/M, z7.s, z11.s\n"
+ "fmla z31.s, p3/M, z6.s, z11.s\n"
+ "fmla z28.s, p3/M, z1.s, z16.s\n"
+ "ld1w { z18.s }, p2/Z, [x7]\n"
+ "fmla z29.s, p3/M, z0.s, z16.s\n"
+ "ld1w { z17.s }, p2/Z, [x7, x13, LSL #2]\n"
+ "fmla z20.s, p3/M, z7.s, z19.s\n"
"addvl x7, x7, #1\n"
- "fmla z30.s, p3/M, z5.s, z13.s\n"
- "fmla z23.s, p3/M, z6.s, z12.s\n"
+ "fmla z22.s, p3/M, z5.s, z19.s\n"
+ "fmla z27.s, p3/M, z6.s, z18.s\n"
"ld1w { z9.s }, p1/Z, [x7, x8, LSL #2]\n"
- "fmla z29.s, p3/M, z0.s, z12.s\n"
- "fmla z31.s, p3/M, z2.s, z11.s\n"
- "fmla z27.s, p3/M, z8.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x8, LSL #2]\n"
- "fmla z26.s, p3/M, z3.s, z12.s\n"
+ "fmla z21.s, p3/M, z0.s, z18.s\n"
+ "fmla z23.s, p3/M, z2.s, z17.s\n"
+ "fmla z31.s, p3/M, z8.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x14, x8, LSL #2]\n"
+ "fmla z30.s, p3/M, z3.s, z18.s\n"
"whilelt p2.s, x21, %x[n_channels]\n"
- "fmla z25.s, p3/M, z8.s, z11.s\n"
- "fmla z28.s, p3/M, z5.s, z11.s\n"
+ "fmla z29.s, p3/M, z8.s, z17.s\n"
+ "fmla z20.s, p3/M, z5.s, z17.s\n"
".inst 0xa040c220 // ld1w { z0.s-z3.s }, pn8.b/Z, [x17]\n"
"addvl x17, x17, #4\n"
- "fmla z29.s, p3/M, z8.s, z13.s\n"
- "fmla z30.s, p3/M, z7.s, z13.s\n"
+ "fmla z21.s, p3/M, z8.s, z16.s\n"
+ "fmla z22.s, p3/M, z7.s, z16.s\n"
"addvl x14, x14, #1\n"
"cmp x25, %x[n_channels]\n"
- "fmla z31.s, p3/M, z6.s, z13.s\n"
- "fmax z23.s, p3/M, z23.s, z17.s\n"
+ "fmla z23.s, p3/M, z6.s, z16.s\n"
+ "fmax z27.s, p3/M, z27.s, z26.s\n"
".inst 0xa040c224 // ld1w { z4.s-z7.s }, pn8.b/Z, [x17]\n"
"addvl x17, x17, #4\n"
- "fmin z23.s, p3/M, z23.s, z16.s\n"
- ".inst 0xc1b0ca38 // fclamp { z24.s-z27.s }, z17.s, z16.s\n"
+ "fmin z27.s, p3/M, z27.s, z14.s\n"
+ ".inst 0xc1aecb5c // fclamp { z28.s-z31.s }, z26.s, z14.s\n"
"ld1w { z11.s }, p1/Z, [x5, x13, LSL #2]\n"
- ".inst 0xc1b0ca3c // fclamp { z28.s-z31.s }, z17.s, z16.s\n"
+ ".inst 0xc1aecb54 // fclamp { z20.s-z23.s }, z26.s, z14.s\n"
"ld1w { z12.s }, p1/Z, [x14]\n"
- "st1w { z23.s }, p0, [x26]\n"
+ "st1w { z27.s }, p0, [x26]\n"
"ld1w { z13.s }, p1/Z, [x6, x8, LSL #2]\n"
- "st1w { z24.s }, p0, [x26, x27, LSL #2]\n"
- "st1w { z25.s }, p0, [x26, x22, LSL #2]\n"
+ "st1w { z28.s }, p0, [x26, x27, LSL #2]\n"
+ "st1w { z29.s }, p0, [x26, x22, LSL #2]\n"
"addvl x26, x26, #1\n"
"ld1w { z8.s }, p3/Z, [x17]\n"
"addvl x17, x17, #1\n"
- "st1w { z26.s }, p0, [x24]\n"
- "st1w { z27.s }, p0, [x24, x27, LSL #2]\n"
- "st1w { z28.s }, p0, [x24, x22, LSL #2]\n"
+ "st1w { z30.s }, p0, [x24]\n"
+ "st1w { z31.s }, p0, [x24, x27, LSL #2]\n"
+ "st1w { z20.s }, p0, [x24, x22, LSL #2]\n"
"addvl x24, x24, #1\n"
- "st1w { z29.s }, p0, [x23]\n"
- "st1w { z30.s }, p0, [x23, x27, LSL #2]\n"
- "st1w { z31.s }, p0, [x23, x22, LSL #2]\n"
+ "st1w { z21.s }, p0, [x23]\n"
+ "st1w { z22.s }, p0, [x23, x27, LSL #2]\n"
+ "st1w { z23.s }, p0, [x23, x22, LSL #2]\n"
"addvl x23, x23, #1\n"
"blt 3b\n"
"4:" // Tile loop: Channel tail
- "movprfx z24, z18\n fmla z24.s, p3/M, z7.s, z9.s\n"
- "movprfx z23, z18\n fmla z23.s, p3/M, z8.s, z9.s\n"
+ "movprfx z28, z24\n fmla z28.s, p3/M, z7.s, z9.s\n"
+ "movprfx z25, z24\n fmla z25.s, p3/M, z8.s, z9.s\n"
"ldr x3, [%x[params_struct], %[offsetof_args_tile_j]]\n"
"add x3, x3, #0x1\n"
- "movprfx z25, z18\n fmla z25.s, p3/M, z6.s, z9.s\n"
- "fmla z24.s, p3/M, z4.s, z13.s\n"
+ "movprfx z29, z24\n fmla z29.s, p3/M, z6.s, z9.s\n"
+ "fmla z28.s, p3/M, z4.s, z13.s\n"
"ldr x2, [%x[params_struct], %[offsetof_args_tile_i]]\n"
"add x21, x2, #0x1\n"
- "movprfx z26, z18\n fmla z26.s, p3/M, z5.s, z9.s\n"
- "movprfx z27, z18\n fmla z27.s, p3/M, z4.s, z9.s\n"
+ "movprfx z30, z24\n fmla z30.s, p3/M, z5.s, z9.s\n"
+ "movprfx z31, z24\n fmla z31.s, p3/M, z4.s, z9.s\n"
"ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
"cmp x3, x20\n"
- "movprfx z28, z18\n fmla z28.s, p3/M, z3.s, z9.s\n"
- "fmla z23.s, p3/M, z0.s, z10.s\n"
- "ld1w { z10.s }, p2/Z, [x7, x15, LSL #2]\n"
+ "movprfx z20, z24\n fmla z20.s, p3/M, z3.s, z9.s\n"
+ "fmla z25.s, p3/M, z0.s, z10.s\n"
+ "ld1w { z27.s }, p2/Z, [x7, x15, LSL #2]\n"
"ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
- "fmla z25.s, p3/M, z2.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x7, x4, LSL #2]\n"
- "movprfx z29, z18\n fmla z29.s, p3/M, z2.s, z9.s\n"
+ "fmla z29.s, p3/M, z2.s, z11.s\n"
+ "ld1w { z17.s }, p2/Z, [x7, x4, LSL #2]\n"
+ "movprfx z21, z24\n fmla z21.s, p3/M, z2.s, z9.s\n"
"csel x2, x2, x21, LT\n"
- "fmla z24.s, p3/M, z6.s, z11.s\n"
- "movprfx z31, z18\n fmla z31.s, p3/M, z0.s, z9.s\n"
+ "fmla z28.s, p3/M, z6.s, z17.s\n"
+ "movprfx z23, z24\n fmla z23.s, p3/M, z0.s, z9.s\n"
"mov p0.b, p2.b\n"
"csel x3, x3, XZR, LT\n"
- "fmla z23.s, p3/M, z5.s, z13.s\n"
- "fmla z25.s, p3/M, z3.s, z13.s\n"
- "cmp x2, x20\n"
- "fmla z26.s, p3/M, z2.s, z13.s\n"
- "fmla z27.s, p3/M, z1.s, z13.s\n"
- "fmla z28.s, p3/M, z0.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x5, x4, LSL #2]\n"
- "fmla z29.s, p3/M, z6.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x14, x13, LSL #2]\n"
- "movprfx z30, z18\n fmla z30.s, p3/M, z1.s, z9.s\n"
- "fmla z24.s, p3/M, z0.s, z13.s\n"
- "fmla z31.s, p3/M, z8.s, z12.s\n"
- "fmla z23.s, p3/M, z7.s, z11.s\n"
- "ld1w { z12.s }, p2/Z, [x5, x15, LSL #2]\n"
- "fmla z30.s, p3/M, z0.s, z11.s\n"
- "fmla z26.s, p3/M, z4.s, z11.s\n"
- "fmla z27.s, p3/M, z3.s, z11.s\n"
- "fmla z29.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x6]\n"
- "fmla z24.s, p3/M, z2.s, z12.s\n"
- "fmla z25.s, p3/M, z1.s, z12.s\n"
- "ld1w { z12.s }, p2/Z, [x16]\n"
- "fmla z28.s, p3/M, z4.s, z10.s\n"
- "fmla z23.s, p3/M, z1.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x6, x13, LSL #2]\n"
- "fmla z30.s, p3/M, z2.s, z10.s\n"
- "fmla z31.s, p3/M, z1.s, z10.s\n"
- "fmla z24.s, p3/M, z8.s, z10.s\n"
- "fmla z25.s, p3/M, z7.s, z10.s\n"
- "fmla z27.s, p3/M, z5.s, z10.s\n"
- "fmla z26.s, p3/M, z0.s, z11.s\n"
- "ld1w { z10.s }, p2/Z, [x16, x8, LSL #2]\n"
- "fmla z28.s, p3/M, z2.s, z13.s\n"
- "fmla z29.s, p3/M, z3.s, z12.s\n"
- "fmla z30.s, p3/M, z4.s, z10.s\n"
- "fmla z31.s, p3/M, z3.s, z10.s\n"
- "fmla z23.s, p3/M, z3.s, z11.s\n"
"fmla z25.s, p3/M, z5.s, z13.s\n"
- "ld1w { z11.s }, p2/Z, [x16, x13, LSL #2]\n"
- "fmla z26.s, p3/M, z6.s, z12.s\n"
- "fmla z27.s, p3/M, z7.s, z10.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x4, LSL #2]\n"
- "fmla z28.s, p3/M, z6.s, z10.s\n"
- "fmla z29.s, p3/M, z5.s, z10.s\n"
- "ld1w { z12.s }, p2/Z, [x6, x4, LSL #2]\n"
- "fmla z31.s, p3/M, z5.s, z11.s\n"
- "fmla z30.s, p3/M, z6.s, z13.s\n"
- "fmla z26.s, p3/M, z8.s, z10.s\n"
- "fmla z28.s, p3/M, z8.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x6, x15, LSL #2]\n"
- "fmla z29.s, p3/M, z7.s, z13.s\n"
- "fmla z24.s, p3/M, z3.s, z12.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x15, LSL #2]\n"
- "fmla z27.s, p3/M, z0.s, z12.s\n"
- "fmla z23.s, p3/M, z4.s, z12.s\n"
- "fmla z30.s, p3/M, z8.s, z13.s\n"
- "fmla z31.s, p3/M, z7.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x16, x15, LSL #2]\n"
- "fmla z26.s, p3/M, z1.s, z12.s\n"
- "fmla z24.s, p3/M, z5.s, z11.s\n"
- "ld1w { z12.s }, p2/Z, [x16, x4, LSL #2]\n"
- "fmla z25.s, p3/M, z4.s, z11.s\n"
- "fmla z27.s, p3/M, z2.s, z11.s\n"
- "fmla z28.s, p3/M, z1.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x5, x8, LSL #2]\n"
- "fmla z29.s, p3/M, z4.s, z12.s\n"
- "fmla z30.s, p3/M, z3.s, z12.s\n"
- "fmla z23.s, p3/M, z2.s, z11.s\n"
- "fmla z31.s, p3/M, z4.s, z13.s\n"
- "fmla z26.s, p3/M, z7.s, z12.s\n"
- "fmla z27.s, p3/M, z6.s, z12.s\n"
- "fmla z24.s, p3/M, z1.s, z11.s\n"
- "ld1w { z12.s }, p2/Z, [x7]\n"
- "fmla z25.s, p3/M, z0.s, z11.s\n"
- "ld1w { z11.s }, p2/Z, [x7, x13, LSL #2]\n"
- "fmla z28.s, p3/M, z7.s, z13.s\n"
- "fmla z30.s, p3/M, z5.s, z13.s\n"
- "fmla z23.s, p3/M, z6.s, z12.s\n"
- "fmla z29.s, p3/M, z0.s, z12.s\n"
- "fmla z31.s, p3/M, z2.s, z11.s\n"
- "fmla z27.s, p3/M, z8.s, z13.s\n"
- "ld1w { z13.s }, p2/Z, [x14, x8, LSL #2]\n"
- "fmla z26.s, p3/M, z3.s, z12.s\n"
- "fmla z25.s, p3/M, z8.s, z11.s\n"
- "fmla z28.s, p3/M, z5.s, z11.s\n"
- "fmla z29.s, p3/M, z8.s, z13.s\n"
- "fmla z30.s, p3/M, z7.s, z13.s\n"
- "fmla z31.s, p3/M, z6.s, z13.s\n"
- "fmax z23.s, p3/M, z23.s, z17.s\n"
- "fmin z23.s, p3/M, z23.s, z16.s\n"
- ".inst 0xc1b0ca38 // fclamp { z24.s-z27.s }, z17.s, z16.s\n"
- "st1w { z23.s }, p0, [x26]\n"
- ".inst 0xc1b0ca3c // fclamp { z28.s-z31.s }, z17.s, z16.s\n"
- "st1w { z24.s }, p0, [x26, x27, LSL #2]\n"
- "st1w { z25.s }, p0, [x26, x22, LSL #2]\n"
- "st1w { z26.s }, p0, [x24]\n"
- "st1w { z27.s }, p0, [x24, x27, LSL #2]\n"
- "st1w { z28.s }, p0, [x24, x22, LSL #2]\n"
- "st1w { z29.s }, p0, [x23]\n"
- "st1w { z30.s }, p0, [x23, x27, LSL #2]\n"
- "st1w { z31.s }, p0, [x23, x22, LSL #2]\n"
+ "fmla z29.s, p3/M, z3.s, z13.s\n"
+ "cmp x2, x20\n"
+ "fmla z30.s, p3/M, z2.s, z13.s\n"
+ "fmla z31.s, p3/M, z1.s, z13.s\n"
+ "fmla z20.s, p3/M, z0.s, z13.s\n"
+ "ld1w { z19.s }, p2/Z, [x5, x4, LSL #2]\n"
+ "fmla z21.s, p3/M, z6.s, z12.s\n"
+ "ld1w { z16.s }, p2/Z, [x14, x13, LSL #2]\n"
+ "movprfx z22, z24\n fmla z22.s, p3/M, z1.s, z9.s\n"
+ "fmla z28.s, p3/M, z0.s, z19.s\n"
+ "fmla z23.s, p3/M, z8.s, z16.s\n"
+ "fmla z25.s, p3/M, z7.s, z17.s\n"
+ "ld1w { z16.s }, p2/Z, [x5, x15, LSL #2]\n"
+ "fmla z22.s, p3/M, z0.s, z17.s\n"
+ "fmla z30.s, p3/M, z4.s, z17.s\n"
+ "fmla z31.s, p3/M, z3.s, z17.s\n"
+ "fmla z21.s, p3/M, z1.s, z17.s\n"
+ "ld1w { z18.s }, p2/Z, [x6]\n"
+ "fmla z28.s, p3/M, z2.s, z16.s\n"
+ "fmla z29.s, p3/M, z1.s, z16.s\n"
+ "ld1w { z17.s }, p2/Z, [x16]\n"
+ "fmla z20.s, p3/M, z4.s, z27.s\n"
+ "fmla z25.s, p3/M, z1.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x6, x13, LSL #2]\n"
+ "fmla z22.s, p3/M, z2.s, z27.s\n"
+ "fmla z23.s, p3/M, z1.s, z27.s\n"
+ "fmla z28.s, p3/M, z8.s, z27.s\n"
+ "fmla z29.s, p3/M, z7.s, z27.s\n"
+ "fmla z31.s, p3/M, z5.s, z27.s\n"
+ "fmla z30.s, p3/M, z0.s, z18.s\n"
+ "ld1w { z19.s }, p2/Z, [x16, x8, LSL #2]\n"
+ "fmla z20.s, p3/M, z2.s, z16.s\n"
+ "fmla z21.s, p3/M, z3.s, z17.s\n"
+ "fmla z22.s, p3/M, z4.s, z19.s\n"
+ "fmla z23.s, p3/M, z3.s, z19.s\n"
+ "fmla z25.s, p3/M, z3.s, z18.s\n"
+ "fmla z29.s, p3/M, z5.s, z16.s\n"
+ "ld1w { z18.s }, p2/Z, [x16, x13, LSL #2]\n"
+ "fmla z30.s, p3/M, z6.s, z17.s\n"
+ "fmla z31.s, p3/M, z7.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x14, x4, LSL #2]\n"
+ "fmla z20.s, p3/M, z6.s, z19.s\n"
+ "fmla z21.s, p3/M, z5.s, z19.s\n"
+ "ld1w { z17.s }, p2/Z, [x6, x4, LSL #2]\n"
+ "fmla z23.s, p3/M, z5.s, z18.s\n"
+ "fmla z22.s, p3/M, z6.s, z16.s\n"
+ "fmla z30.s, p3/M, z8.s, z19.s\n"
+ "fmla z20.s, p3/M, z8.s, z18.s\n"
+ "ld1w { z18.s }, p2/Z, [x6, x15, LSL #2]\n"
+ "fmla z21.s, p3/M, z7.s, z16.s\n"
+ "fmla z28.s, p3/M, z3.s, z17.s\n"
+ "ld1w { z16.s }, p2/Z, [x14, x15, LSL #2]\n"
+ "fmla z31.s, p3/M, z0.s, z17.s\n"
+ "fmla z25.s, p3/M, z4.s, z17.s\n"
+ "fmla z22.s, p3/M, z8.s, z16.s\n"
+ "fmla z23.s, p3/M, z7.s, z16.s\n"
+ "ld1w { z19.s }, p2/Z, [x16, x15, LSL #2]\n"
+ "fmla z30.s, p3/M, z1.s, z17.s\n"
+ "fmla z28.s, p3/M, z5.s, z18.s\n"
+ "ld1w { z17.s }, p2/Z, [x16, x4, LSL #2]\n"
+ "fmla z29.s, p3/M, z4.s, z18.s\n"
+ "fmla z31.s, p3/M, z2.s, z18.s\n"
+ "fmla z20.s, p3/M, z1.s, z18.s\n"
+ "ld1w { z16.s }, p2/Z, [x5, x8, LSL #2]\n"
+ "fmla z21.s, p3/M, z4.s, z17.s\n"
+ "fmla z22.s, p3/M, z3.s, z17.s\n"
+ "fmla z25.s, p3/M, z2.s, z16.s\n"
+ "fmla z23.s, p3/M, z4.s, z19.s\n"
+ "fmla z30.s, p3/M, z7.s, z17.s\n"
+ "fmla z31.s, p3/M, z6.s, z17.s\n"
+ "fmla z28.s, p3/M, z1.s, z16.s\n"
+ "ld1w { z18.s }, p2/Z, [x7]\n"
+ "fmla z29.s, p3/M, z0.s, z16.s\n"
+ "ld1w { z17.s }, p2/Z, [x7, x13, LSL #2]\n"
+ "fmla z20.s, p3/M, z7.s, z19.s\n"
+ "fmla z22.s, p3/M, z5.s, z19.s\n"
+ "fmla z25.s, p3/M, z6.s, z18.s\n"
+ "fmla z21.s, p3/M, z0.s, z18.s\n"
+ "fmla z23.s, p3/M, z2.s, z17.s\n"
+ "fmla z31.s, p3/M, z8.s, z19.s\n"
+ "ld1w { z16.s }, p2/Z, [x14, x8, LSL #2]\n"
+ "fmla z30.s, p3/M, z3.s, z18.s\n"
+ "fmla z29.s, p3/M, z8.s, z17.s\n"
+ "fmla z20.s, p3/M, z5.s, z17.s\n"
+ "fmla z21.s, p3/M, z8.s, z16.s\n"
+ "fmla z22.s, p3/M, z7.s, z16.s\n"
+ "fmla z23.s, p3/M, z6.s, z16.s\n"
+ "fmax z25.s, p3/M, z25.s, z26.s\n"
+ "fmin z25.s, p3/M, z25.s, z14.s\n"
+ ".inst 0xc1aecb5c // fclamp { z28.s-z31.s }, z26.s, z14.s\n"
+ "st1w { z25.s }, p0, [x26]\n"
+ ".inst 0xc1aecb54 // fclamp { z20.s-z23.s }, z26.s, z14.s\n"
+ "st1w { z28.s }, p0, [x26, x27, LSL #2]\n"
+ "st1w { z29.s }, p0, [x26, x22, LSL #2]\n"
+ "st1w { z30.s }, p0, [x24]\n"
+ "st1w { z31.s }, p0, [x24, x27, LSL #2]\n"
+ "st1w { z20.s }, p0, [x24, x22, LSL #2]\n"
+ "st1w { z21.s }, p0, [x23]\n"
+ "st1w { z22.s }, p0, [x23, x27, LSL #2]\n"
+ "st1w { z23.s }, p0, [x23, x22, LSL #2]\n"
"blt 1b\n"
".inst 0xd503467f // SMSTOP\n"
: