aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp532
1 files changed, 267 insertions, 265 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
index 027cc9e5a2..61cec2b66d 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
@@ -22,12 +22,13 @@
* SOFTWARE.
*/
-#if defined(__aarch64__)
#include "arm_gemm.hpp"
#include <cstddef>
#include <cstdint>
+#if defined(__aarch64__)
+
namespace arm_conv {
namespace depthwise {
@@ -42,133 +43,133 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
__asm__ __volatile__(
"ldr q12, [%x[params], #0x0]\n"
"ldr q8, [%x[params], #0x10]\n"
- "movi v28.16b, #0x1\n"
- "movi v18.4s, #0x0\n"
+ "movi v30.16b, #0x1\n"
+ "movi v17.4s, #0x0\n"
"ldr q9, [%x[params], #0x20]\n"
"ldr q10, [%x[params], #0x30]\n"
- "movi v31.4s, #0x0\n"
- "movi v24.4s, #0x0\n"
+ "movi v16.4s, #0x0\n"
+ "movi v25.4s, #0x0\n"
"ldr q11, [%x[params], #0x40]\n"
"ldr x20, [%x[inptrs], #0x18]\n"
- "movi v30.4s, #0x0\n"
- "movi v21.4s, #0x0\n"
+ "movi v24.4s, #0x0\n"
+ "movi v31.4s, #0x0\n"
"ld1 { v3.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x20]\n"
- "mov v16.16b, v3.16b\n"
- "ext v16.16b, v16.16b, v16.16b, #0x1\n"
+ "mov v26.16b, v3.16b\n"
+ "ext v26.16b, v26.16b, v26.16b, #0x1\n"
"ld1 { v4.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x10]\n"
- "mov v15.16b, v4.16b\n"
- "ext v15.16b, v15.16b, v15.16b, #0x1\n"
+ "mov v21.16b, v4.16b\n"
+ "ext v21.16b, v21.16b, v21.16b, #0x1\n"
"ld1 { v2.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x8]\n"
- "mov v20.16b, v2.16b\n"
- "ext v20.16b, v20.16b, v20.16b, #0x1\n"
+ "mov v27.16b, v2.16b\n"
+ "ext v27.16b, v27.16b, v27.16b, #0x1\n"
"ld1 { v1.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x28]\n"
- "zip1 v3.2d, v3.2d, v16.2d\n"
- "zip1 v4.2d, v4.2d, v15.2d\n"
+ "zip1 v3.2d, v3.2d, v26.2d\n"
+ "zip1 v4.2d, v4.2d, v21.2d\n"
"ld1 { v5.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x30]\n"
"mov v26.16b, v1.16b\n"
- "mov v13.16b, v5.16b\n"
+ "mov v22.16b, v5.16b\n"
"ld1 { v6.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x38]\n"
"mov v19.16b, v6.16b\n"
"ext v26.16b, v26.16b, v26.16b, #0x1\n"
"ld1 { v7.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x0]\n"
- "mov v17.16b, v7.16b\n"
- "zip1 v2.2d, v2.2d, v20.2d\n"
+ "mov v21.16b, v7.16b\n"
+ "zip1 v2.2d, v2.2d, v27.2d\n"
"ld1 { v0.16b }, [x20]\n"
- "ext v13.16b, v13.16b, v13.16b, #0x1\n"
+ "ext v22.16b, v22.16b, v22.16b, #0x1\n"
"ext v19.16b, v19.16b, v19.16b, #0x1\n"
- ".inst 0x6f83e392 // udot v18.4s, v28.16b, v3.4b[0]\n"
- "ext v17.16b, v17.16b, v17.16b, #0x1\n"
- ".inst 0x6f83eb9f // udot v31.4s, v28.16b, v3.4b[2]\n"
- ".inst 0x6f84e398 // udot v24.4s, v28.16b, v4.4b[0]\n"
+ ".inst 0x6f83e3d1 // udot v17.4s, v30.16b, v3.4b[0]\n"
+ "ext v21.16b, v21.16b, v21.16b, #0x1\n"
+ ".inst 0x6f83ebd0 // udot v16.4s, v30.16b, v3.4b[2]\n"
+ ".inst 0x6f84e3d9 // udot v25.4s, v30.16b, v4.4b[0]\n"
"add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n"
"ld1r { v23.4s }, [x20]\n"
- ".inst 0x6f84eb9e // udot v30.4s, v28.16b, v4.4b[2]\n"
- "mov v16.16b, v0.16b\n"
- ".inst 0x6f82e395 // udot v21.4s, v28.16b, v2.4b[0]\n"
- "movi v20.4s, #0x0\n"
- "movi v29.4s, #0x1\n"
- ".inst 0x6f82eb94 // udot v20.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x6f84ebd8 // udot v24.4s, v30.16b, v4.4b[2]\n"
+ "mov v18.16b, v0.16b\n"
+ ".inst 0x6f82e3df // udot v31.4s, v30.16b, v2.4b[0]\n"
+ "movi v29.4s, #0x0\n"
+ "movi v28.4s, #0x1\n"
+ ".inst 0x6f82ebdd // udot v29.4s, v30.16b, v2.4b[2]\n"
"add x20, %x[qp], %[offsetof_Requantize32_c_offset]\n"
- "ld1r { v14.4s }, [x20]\n"
- "ext v16.16b, v16.16b, v16.16b, #0x1\n"
+ "ld1r { v13.4s }, [x20]\n"
+ "ext v18.16b, v18.16b, v18.16b, #0x1\n"
"zip1 v1.2d, v1.2d, v26.2d\n"
- ".inst 0x6fa3e3b2 // udot v18.4s, v29.16b, v3.4b[1]\n"
- "zip1 v5.2d, v5.2d, v13.2d\n"
+ ".inst 0x6fa3e391 // udot v17.4s, v28.16b, v3.4b[1]\n"
+ "zip1 v5.2d, v5.2d, v22.2d\n"
"zip1 v6.2d, v6.2d, v19.2d\n"
- ".inst 0x6fa3ebbf // udot v31.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x6fa3eb90 // udot v16.4s, v28.16b, v3.4b[3]\n"
"add x20, %x[qp], %[offsetof_Requantize32_minval]\n"
- "ld1r { v13.4s }, [x20]\n"
- "zip1 v7.2d, v7.2d, v17.2d\n"
+ "ld1r { v14.4s }, [x20]\n"
+ "zip1 v7.2d, v7.2d, v21.2d\n"
"movi v22.4s, #0x0\n"
- ".inst 0x6fa4e3b8 // udot v24.4s, v29.16b, v4.4b[1]\n"
- "movi v26.4s, #0x0\n"
- ".inst 0x6fa4ebbe // udot v30.4s, v29.16b, v4.4b[3]\n"
- ".inst 0x6f81e396 // udot v22.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x6fa4e399 // udot v25.4s, v28.16b, v4.4b[1]\n"
+ "movi v21.4s, #0x0\n"
+ ".inst 0x6fa4eb98 // udot v24.4s, v28.16b, v4.4b[3]\n"
+ ".inst 0x6f81e3d6 // udot v22.4s, v30.16b, v1.4b[0]\n"
"add x20, %x[qp], %[offsetof_Requantize32_maxval]\n"
"ld1r { v15.4s }, [x20]\n"
- "movi v25.4s, #0x0\n"
+ "movi v26.4s, #0x0\n"
"movi v27.4s, #0x0\n"
- ".inst 0x6f81eb9a // udot v26.4s, v28.16b, v1.4b[2]\n"
- "zip1 v0.2d, v0.2d, v16.2d\n"
+ ".inst 0x6f81ebd5 // udot v21.4s, v30.16b, v1.4b[2]\n"
+ "movi v20.4s, #0x0\n"
"movi v19.4s, #0x0\n"
- ".inst 0x6f85e399 // udot v25.4s, v28.16b, v5.4b[0]\n"
+ ".inst 0x6f85e3da // udot v26.4s, v30.16b, v5.4b[0]\n"
"cmp %x[n_channels], #0x4\n"
- ".inst 0x6f85eb9b // udot v27.4s, v28.16b, v5.4b[2]\n"
- ".inst 0x6f86e393 // udot v19.4s, v28.16b, v6.4b[0]\n"
- "add v24.4s, v18.4s, v24.4s\n"
- "mov x9, #0x0\n"
+ "zip1 v0.2d, v0.2d, v18.2d\n"
"movi v18.4s, #0x0\n"
- ".inst 0x6f86eb92 // udot v18.4s, v28.16b, v6.4b[2]\n"
- ".inst 0x6fa2e3b5 // udot v21.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x6f85ebdb // udot v27.4s, v30.16b, v5.4b[2]\n"
+ "mov x9, #0x0\n"
+ ".inst 0x6f86e3d4 // udot v20.4s, v30.16b, v6.4b[0]\n"
+ ".inst 0x6f86ebd3 // udot v19.4s, v30.16b, v6.4b[2]\n"
+ "add v17.4s, v17.4s, v25.4s\n"
"mov x28, #0x0\n"
- ".inst 0x6fa2ebb4 // udot v20.4s, v29.16b, v2.4b[3]\n"
- "add v17.4s, v31.4s, v30.4s\n"
- ".inst 0x6fa1e3b6 // udot v22.4s, v29.16b, v1.4b[1]\n"
+ "movi v25.4s, #0x0\n"
+ ".inst 0x6f87e3d2 // udot v18.4s, v30.16b, v7.4b[0]\n"
+ ".inst 0x6f87ebd9 // udot v25.4s, v30.16b, v7.4b[2]\n"
"ldp x27, x26, [%x[outptrs], #0x0]\n"
- "movi v16.4s, #0x0\n"
- ".inst 0x6f87e390 // udot v16.4s, v28.16b, v7.4b[0]\n"
- ".inst 0x6fa1ebba // udot v26.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x6fa2e39f // udot v31.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x6fa2eb9d // udot v29.4s, v28.16b, v2.4b[3]\n"
+ "add v16.4s, v16.4s, v24.4s\n"
"ldp x25, x24, [%x[outptrs], #0x10]\n"
- ".inst 0x6fa5e3b9 // udot v25.4s, v29.16b, v5.4b[1]\n"
- ".inst 0x6fa5ebbb // udot v27.4s, v29.16b, v5.4b[3]\n"
- "add v30.4s, v21.4s, v24.4s\n"
+ "movi v24.4s, #0x0\n"
+ ".inst 0x6f80e3d8 // udot v24.4s, v30.16b, v0.4b[0]\n"
+ ".inst 0x6fa1e396 // udot v22.4s, v28.16b, v1.4b[1]\n"
"ldp x23, x22, [%x[outptrs], #0x20]\n"
- ".inst 0x6fa6e3b3 // udot v19.4s, v29.16b, v6.4b[1]\n"
- ".inst 0x6fa6ebb2 // udot v18.4s, v29.16b, v6.4b[3]\n"
- "add v31.4s, v20.4s, v17.4s\n"
+ ".inst 0x6fa1eb95 // udot v21.4s, v28.16b, v1.4b[3]\n"
+ ".inst 0x6fa5e39a // udot v26.4s, v28.16b, v5.4b[1]\n"
+ "add v31.4s, v31.4s, v17.4s\n"
"ldp x21, x20, [%x[outptrs], #0x30]\n"
- ".inst 0x6fa7e3b0 // udot v16.4s, v29.16b, v7.4b[1]\n"
- "add v22.4s, v22.4s, v30.4s\n"
+ ".inst 0x6fa5eb9b // udot v27.4s, v28.16b, v5.4b[3]\n"
+ ".inst 0x6fa6e394 // udot v20.4s, v28.16b, v6.4b[1]\n"
+ "add v29.4s, v29.4s, v16.4s\n"
"add %x[params], %x[params], #0x50\n"
- "add v21.4s, v26.4s, v31.4s\n"
- "add v20.4s, v25.4s, v19.4s\n"
- "add v19.4s, v27.4s, v18.4s\n"
- "add v18.4s, v16.4s, v24.4s\n"
- "movi v16.4s, #0x0\n"
- ".inst 0x6f87eb90 // udot v16.4s, v28.16b, v7.4b[2]\n"
- ".inst 0x6fa7ebb0 // udot v16.4s, v29.16b, v7.4b[3]\n"
- "add v17.4s, v16.4s, v17.4s\n"
- "movi v16.4s, #0x0\n"
- ".inst 0x6f80e390 // udot v16.4s, v28.16b, v0.4b[0]\n"
- ".inst 0x6fa0e3b0 // udot v16.4s, v29.16b, v0.4b[1]\n"
- "add v24.4s, v22.4s, v16.4s\n"
- "add v26.4s, v22.4s, v25.4s\n"
- "movi v16.4s, #0x0\n"
- ".inst 0x6f80eb90 // udot v16.4s, v28.16b, v0.4b[2]\n"
- ".inst 0x6fa0ebb0 // udot v16.4s, v29.16b, v0.4b[3]\n"
- "add v25.4s, v21.4s, v16.4s\n"
- "add v27.4s, v21.4s, v27.4s\n"
- "add v28.4s, v20.4s, v30.4s\n"
- "add v29.4s, v19.4s, v31.4s\n"
- "add v30.4s, v18.4s, v20.4s\n"
- "add v31.4s, v17.4s, v19.4s\n"
+ ".inst 0x6fa6eb93 // udot v19.4s, v28.16b, v6.4b[3]\n"
+ ".inst 0x6fa7e392 // udot v18.4s, v28.16b, v7.4b[1]\n"
+ "add v22.4s, v22.4s, v31.4s\n"
+ ".inst 0x6fa7eb99 // udot v25.4s, v28.16b, v7.4b[3]\n"
+ ".inst 0x6fa0e398 // udot v24.4s, v28.16b, v0.4b[1]\n"
+ "add v21.4s, v21.4s, v29.4s\n"
+ "add v20.4s, v26.4s, v20.4s\n"
+ "add v19.4s, v27.4s, v19.4s\n"
+ "add v18.4s, v18.4s, v17.4s\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x6f80ebd1 // udot v17.4s, v30.16b, v0.4b[2]\n"
+ ".inst 0x6fa0eb91 // udot v17.4s, v28.16b, v0.4b[3]\n"
+ "add v16.4s, v25.4s, v16.4s\n"
+ "add v24.4s, v22.4s, v24.4s\n"
+ "add v25.4s, v21.4s, v17.4s\n"
+ "add v26.4s, v26.4s, v22.4s\n"
+ "add v27.4s, v27.4s, v21.4s\n"
+ "add v28.4s, v20.4s, v31.4s\n"
+ "add v29.4s, v19.4s, v29.4s\n"
+ "add v30.4s, v20.4s, v18.4s\n"
+ "add v31.4s, v19.4s, v16.4s\n"
"neg v23.4s, v23.4s\n"
"mul v24.4s, v24.4s, v23.4s\n"
"mul v25.4s, v25.4s, v23.4s\n"
@@ -194,11 +195,11 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
"add v31.4s, v31.4s, v12.4s\n"
"ble 2f\n"
"1:" // Loop
- "ldr q21, [%x[params], #0x60]\n"
- "ldr q20, [%x[params], #0x70]\n"
+ "ldr q12, [%x[params], #0x60]\n"
+ "ldr q21, [%x[params], #0x70]\n"
".inst 0x6f80e118 // udot v24.4s, v8.16b, v0.4b[0]\n"
".inst 0x6f80e919 // udot v25.4s, v8.16b, v0.4b[2]\n"
- "ldr q12, [%x[params], #0x80]\n"
+ "ldr q20, [%x[params], #0x80]\n"
".inst 0x6f81e11a // udot v26.4s, v8.16b, v1.4b[0]\n"
".inst 0x6f81e91b // udot v27.4s, v8.16b, v1.4b[2]\n"
"sub %x[n_channels], %x[n_channels], #0x4\n"
@@ -212,7 +213,7 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
".inst 0x6f82e91d // udot v29.4s, v8.16b, v2.4b[2]\n"
".inst 0x6f83e11e // udot v30.4s, v8.16b, v3.4b[0]\n"
".inst 0x6f83e91f // udot v31.4s, v8.16b, v3.4b[2]\n"
- "ldr q8, [%x[params], #0x0]\n"
+ "ldr q17, [%x[params], #0x0]\n"
".inst 0x6f81e158 // udot v24.4s, v10.16b, v1.4b[0]\n"
".inst 0x6f81e959 // udot v25.4s, v10.16b, v1.4b[2]\n"
".inst 0x6f82e15a // udot v26.4s, v10.16b, v2.4b[0]\n"
@@ -221,7 +222,7 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
".inst 0x6fa2e93d // udot v29.4s, v9.16b, v2.4b[3]\n"
".inst 0x6fa3e13e // udot v30.4s, v9.16b, v3.4b[1]\n"
".inst 0x6fa3e93f // udot v31.4s, v9.16b, v3.4b[3]\n"
- "ldr q9, [%x[params], #0x10]\n"
+ "ldr q16, [%x[params], #0x10]\n"
".inst 0x6fa1e178 // udot v24.4s, v11.16b, v1.4b[1]\n"
".inst 0x6fa1e979 // udot v25.4s, v11.16b, v1.4b[3]\n"
".inst 0x6fa2e17a // udot v26.4s, v11.16b, v2.4b[1]\n"
@@ -230,115 +231,115 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
".inst 0x6f83e95d // udot v29.4s, v10.16b, v3.4b[2]\n"
".inst 0x6f84e15e // udot v30.4s, v10.16b, v4.4b[0]\n"
".inst 0x6f84e95f // udot v31.4s, v10.16b, v4.4b[2]\n"
- "ldr q10, [%x[params], #0x20]\n"
- ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- ".inst 0x6f82e919 // udot v25.4s, v8.16b, v2.4b[2]\n"
- ".inst 0x6f83e11a // udot v26.4s, v8.16b, v3.4b[0]\n"
- ".inst 0x6f83e91b // udot v27.4s, v8.16b, v3.4b[2]\n"
+ "ldr q19, [%x[params], #0x20]\n"
+ ".inst 0x6f82e238 // udot v24.4s, v17.16b, v2.4b[0]\n"
+ ".inst 0x6f82ea39 // udot v25.4s, v17.16b, v2.4b[2]\n"
+ ".inst 0x6f83e23a // udot v26.4s, v17.16b, v3.4b[0]\n"
+ ".inst 0x6f83ea3b // udot v27.4s, v17.16b, v3.4b[2]\n"
".inst 0x6fa3e17c // udot v28.4s, v11.16b, v3.4b[1]\n"
".inst 0x6fa3e97d // udot v29.4s, v11.16b, v3.4b[3]\n"
".inst 0x6fa4e17e // udot v30.4s, v11.16b, v4.4b[1]\n"
".inst 0x6fa4e97f // udot v31.4s, v11.16b, v4.4b[3]\n"
- "ldr q11, [%x[params], #0x30]\n"
- ".inst 0x6fa2e138 // udot v24.4s, v9.16b, v2.4b[1]\n"
- ".inst 0x6fa2e939 // udot v25.4s, v9.16b, v2.4b[3]\n"
- ".inst 0x6fa3e13a // udot v26.4s, v9.16b, v3.4b[1]\n"
- ".inst 0x6fa3e93b // udot v27.4s, v9.16b, v3.4b[3]\n"
- ".inst 0x6f84e11c // udot v28.4s, v8.16b, v4.4b[0]\n"
- ".inst 0x6f84e91d // udot v29.4s, v8.16b, v4.4b[2]\n"
- ".inst 0x6f85e11e // udot v30.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x6f85e91f // udot v31.4s, v8.16b, v5.4b[2]\n"
- "ldr q8, [%x[params], #0x40]\n"
- ".inst 0x6f83e158 // udot v24.4s, v10.16b, v3.4b[0]\n"
- ".inst 0x6f83e959 // udot v25.4s, v10.16b, v3.4b[2]\n"
- ".inst 0x6f84e15a // udot v26.4s, v10.16b, v4.4b[0]\n"
- ".inst 0x6f84e95b // udot v27.4s, v10.16b, v4.4b[2]\n"
- ".inst 0x6fa4e13c // udot v28.4s, v9.16b, v4.4b[1]\n"
- ".inst 0x6fa4e93d // udot v29.4s, v9.16b, v4.4b[3]\n"
- ".inst 0x6fa5e13e // udot v30.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x6fa5e93f // udot v31.4s, v9.16b, v5.4b[3]\n"
- "ldr q9, [%x[params], #0x50]\n"
- ".inst 0x6fa3e178 // udot v24.4s, v11.16b, v3.4b[1]\n"
- ".inst 0x6fa3e979 // udot v25.4s, v11.16b, v3.4b[3]\n"
- ".inst 0x6fa4e17a // udot v26.4s, v11.16b, v4.4b[1]\n"
- ".inst 0x6fa4e97b // udot v27.4s, v11.16b, v4.4b[3]\n"
- ".inst 0x6f85e15c // udot v28.4s, v10.16b, v5.4b[0]\n"
- ".inst 0x6f85e95d // udot v29.4s, v10.16b, v5.4b[2]\n"
- ".inst 0x6f86e15e // udot v30.4s, v10.16b, v6.4b[0]\n"
- ".inst 0x6f86e95f // udot v31.4s, v10.16b, v6.4b[2]\n"
+ "ldr q18, [%x[params], #0x30]\n"
+ ".inst 0x6fa2e218 // udot v24.4s, v16.16b, v2.4b[1]\n"
+ ".inst 0x6fa2ea19 // udot v25.4s, v16.16b, v2.4b[3]\n"
+ ".inst 0x6fa3e21a // udot v26.4s, v16.16b, v3.4b[1]\n"
+ ".inst 0x6fa3ea1b // udot v27.4s, v16.16b, v3.4b[3]\n"
+ ".inst 0x6f84e23c // udot v28.4s, v17.16b, v4.4b[0]\n"
+ ".inst 0x6f84ea3d // udot v29.4s, v17.16b, v4.4b[2]\n"
+ ".inst 0x6f85e23e // udot v30.4s, v17.16b, v5.4b[0]\n"
+ ".inst 0x6f85ea3f // udot v31.4s, v17.16b, v5.4b[2]\n"
+ "ldr q17, [%x[params], #0x40]\n"
+ ".inst 0x6f83e278 // udot v24.4s, v19.16b, v3.4b[0]\n"
+ ".inst 0x6f83ea79 // udot v25.4s, v19.16b, v3.4b[2]\n"
+ ".inst 0x6f84e27a // udot v26.4s, v19.16b, v4.4b[0]\n"
+ ".inst 0x6f84ea7b // udot v27.4s, v19.16b, v4.4b[2]\n"
+ ".inst 0x6fa4e21c // udot v28.4s, v16.16b, v4.4b[1]\n"
+ ".inst 0x6fa4ea1d // udot v29.4s, v16.16b, v4.4b[3]\n"
+ ".inst 0x6fa5e21e // udot v30.4s, v16.16b, v5.4b[1]\n"
+ ".inst 0x6fa5ea1f // udot v31.4s, v16.16b, v5.4b[3]\n"
+ "ldr q16, [%x[params], #0x50]\n"
+ ".inst 0x6fa3e258 // udot v24.4s, v18.16b, v3.4b[1]\n"
+ ".inst 0x6fa3ea59 // udot v25.4s, v18.16b, v3.4b[3]\n"
+ ".inst 0x6fa4e25a // udot v26.4s, v18.16b, v4.4b[1]\n"
+ ".inst 0x6fa4ea5b // udot v27.4s, v18.16b, v4.4b[3]\n"
+ ".inst 0x6f85e27c // udot v28.4s, v19.16b, v5.4b[0]\n"
+ ".inst 0x6f85ea7d // udot v29.4s, v19.16b, v5.4b[2]\n"
+ ".inst 0x6f86e27e // udot v30.4s, v19.16b, v6.4b[0]\n"
+ ".inst 0x6f86ea7f // udot v31.4s, v19.16b, v6.4b[2]\n"
"ldr q10, [%x[params], #0xb0]\n"
- ".inst 0x6f84e118 // udot v24.4s, v8.16b, v4.4b[0]\n"
- ".inst 0x6f84e919 // udot v25.4s, v8.16b, v4.4b[2]\n"
- ".inst 0x6f85e11a // udot v26.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x6f85e91b // udot v27.4s, v8.16b, v5.4b[2]\n"
- ".inst 0x6fa5e17c // udot v28.4s, v11.16b, v5.4b[1]\n"
- ".inst 0x6fa5e97d // udot v29.4s, v11.16b, v5.4b[3]\n"
- ".inst 0x6fa6e17e // udot v30.4s, v11.16b, v6.4b[1]\n"
- ".inst 0x6fa6e97f // udot v31.4s, v11.16b, v6.4b[3]\n"
+ ".inst 0x6f84e238 // udot v24.4s, v17.16b, v4.4b[0]\n"
+ ".inst 0x6f84ea39 // udot v25.4s, v17.16b, v4.4b[2]\n"
+ ".inst 0x6f85e23a // udot v26.4s, v17.16b, v5.4b[0]\n"
+ ".inst 0x6f85ea3b // udot v27.4s, v17.16b, v5.4b[2]\n"
+ ".inst 0x6fa5e25c // udot v28.4s, v18.16b, v5.4b[1]\n"
+ ".inst 0x6fa5ea5d // udot v29.4s, v18.16b, v5.4b[3]\n"
+ ".inst 0x6fa6e25e // udot v30.4s, v18.16b, v6.4b[1]\n"
+ ".inst 0x6fa6ea5f // udot v31.4s, v18.16b, v6.4b[3]\n"
"ldr q11, [%x[params], #0xc0]\n"
- ".inst 0x6fa4e138 // udot v24.4s, v9.16b, v4.4b[1]\n"
- ".inst 0x6fa4e939 // udot v25.4s, v9.16b, v4.4b[3]\n"
- "sqrdmulh v24.4s, v24.4s, v21.4s\n"
- ".inst 0x6fa5e13a // udot v26.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x6fa5e93b // udot v27.4s, v9.16b, v5.4b[3]\n"
- "sqrdmulh v25.4s, v25.4s, v21.4s\n"
- ".inst 0x6f86e11c // udot v28.4s, v8.16b, v6.4b[0]\n"
- ".inst 0x6f86e91d // udot v29.4s, v8.16b, v6.4b[2]\n"
- "sqrdmulh v26.4s, v26.4s, v21.4s\n"
- ".inst 0x6f87e11e // udot v30.4s, v8.16b, v7.4b[0]\n"
- ".inst 0x6f87e91f // udot v31.4s, v8.16b, v7.4b[2]\n"
+ ".inst 0x6fa4e218 // udot v24.4s, v16.16b, v4.4b[1]\n"
+ ".inst 0x6fa4ea19 // udot v25.4s, v16.16b, v4.4b[3]\n"
+ "sqrdmulh v24.4s, v24.4s, v12.4s\n"
+ ".inst 0x6fa5e21a // udot v26.4s, v16.16b, v5.4b[1]\n"
+ ".inst 0x6fa5ea1b // udot v27.4s, v16.16b, v5.4b[3]\n"
+ "sqrdmulh v25.4s, v25.4s, v12.4s\n"
+ ".inst 0x6f86e23c // udot v28.4s, v17.16b, v6.4b[0]\n"
+ ".inst 0x6f86ea3d // udot v29.4s, v17.16b, v6.4b[2]\n"
+ "sqrdmulh v26.4s, v26.4s, v12.4s\n"
+ ".inst 0x6f87e23e // udot v30.4s, v17.16b, v7.4b[0]\n"
+ ".inst 0x6f87ea3f // udot v31.4s, v17.16b, v7.4b[2]\n"
"ldr q8, [%x[params], #0x90]\n"
- "sqrdmulh v27.4s, v27.4s, v21.4s\n"
- ".inst 0x6fa6e13c // udot v28.4s, v9.16b, v6.4b[1]\n"
- ".inst 0x6fa6e93d // udot v29.4s, v9.16b, v6.4b[3]\n"
- "and v19.16b, v24.16b, v20.16b\n"
- ".inst 0x6fa7e13e // udot v30.4s, v9.16b, v7.4b[1]\n"
- ".inst 0x6fa7e93f // udot v31.4s, v9.16b, v7.4b[3]\n"
+ "sqrdmulh v27.4s, v27.4s, v12.4s\n"
+ ".inst 0x6fa6e21c // udot v28.4s, v16.16b, v6.4b[1]\n"
+ ".inst 0x6fa6ea1d // udot v29.4s, v16.16b, v6.4b[3]\n"
+ "and v19.16b, v24.16b, v21.16b\n"
+ ".inst 0x6fa7e21e // udot v30.4s, v16.16b, v7.4b[1]\n"
+ ".inst 0x6fa7ea1f // udot v31.4s, v16.16b, v7.4b[3]\n"
"ldr q9, [%x[params], #0xa0]\n"
- "and v18.16b, v25.16b, v20.16b\n"
+ "and v18.16b, v25.16b, v21.16b\n"
+ "and v17.16b, v26.16b, v21.16b\n"
+ "and v16.16b, v27.16b, v21.16b\n"
+ "add %x[params], %x[params], #0xd0\n"
"sshr v19.4s, v19.4s, #0x1f\n"
"sshr v18.4s, v18.4s, #0x1f\n"
- "add %x[params], %x[params], #0xd0\n"
- "sqrdmulh v28.4s, v28.4s, v21.4s\n"
- "sqrdmulh v29.4s, v29.4s, v21.4s\n"
- "sqrdmulh v30.4s, v30.4s, v21.4s\n"
- "sqrdmulh v31.4s, v31.4s, v21.4s\n"
- "and v17.16b, v26.16b, v20.16b\n"
"sshr v17.4s, v17.4s, #0x1f\n"
- "sqadd v24.4s, v24.4s, v19.4s\n"
- "and v16.16b, v27.16b, v20.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
+ "sqrdmulh v28.4s, v28.4s, v12.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v12.4s\n"
+ "sqrdmulh v30.4s, v30.4s, v12.4s\n"
+ "sqrdmulh v31.4s, v31.4s, v12.4s\n"
+ "sqadd v24.4s, v24.4s, v19.4s\n"
"sqadd v25.4s, v25.4s, v18.4s\n"
"sqadd v26.4s, v26.4s, v17.4s\n"
"sqadd v27.4s, v27.4s, v16.4s\n"
- "and v19.16b, v28.16b, v20.16b\n"
- "and v18.16b, v29.16b, v20.16b\n"
- "and v17.16b, v30.16b, v20.16b\n"
+ "and v19.16b, v28.16b, v21.16b\n"
+ "and v18.16b, v29.16b, v21.16b\n"
+ "and v17.16b, v30.16b, v21.16b\n"
+ "and v16.16b, v31.16b, v21.16b\n"
"sshr v19.4s, v19.4s, #0x1f\n"
"sshr v18.4s, v18.4s, #0x1f\n"
"sshr v17.4s, v17.4s, #0x1f\n"
- "sqadd v28.4s, v28.4s, v19.4s\n"
- "and v16.16b, v31.16b, v20.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
+ "sqadd v28.4s, v28.4s, v19.4s\n"
"sqadd v29.4s, v29.4s, v18.4s\n"
"sqadd v30.4s, v30.4s, v17.4s\n"
"sqadd v31.4s, v31.4s, v16.4s\n"
- "srshl v24.4s, v24.4s, v20.4s\n"
- "srshl v25.4s, v25.4s, v20.4s\n"
- "srshl v26.4s, v26.4s, v20.4s\n"
- "srshl v27.4s, v27.4s, v20.4s\n"
- "srshl v28.4s, v28.4s, v20.4s\n"
- "srshl v29.4s, v29.4s, v20.4s\n"
- "srshl v30.4s, v30.4s, v20.4s\n"
- "srshl v31.4s, v31.4s, v20.4s\n"
- "add v24.4s, v24.4s, v14.4s\n"
- "add v25.4s, v25.4s, v14.4s\n"
- "add v26.4s, v26.4s, v14.4s\n"
- "add v27.4s, v27.4s, v14.4s\n"
- "add v28.4s, v28.4s, v14.4s\n"
- "add v29.4s, v29.4s, v14.4s\n"
- "add v30.4s, v30.4s, v14.4s\n"
- "add v31.4s, v31.4s, v14.4s\n"
+ "srshl v24.4s, v24.4s, v21.4s\n"
+ "srshl v25.4s, v25.4s, v21.4s\n"
+ "srshl v26.4s, v26.4s, v21.4s\n"
+ "srshl v27.4s, v27.4s, v21.4s\n"
+ "srshl v28.4s, v28.4s, v21.4s\n"
+ "srshl v29.4s, v29.4s, v21.4s\n"
+ "srshl v30.4s, v30.4s, v21.4s\n"
+ "srshl v31.4s, v31.4s, v21.4s\n"
+ "add v24.4s, v24.4s, v13.4s\n"
+ "add v25.4s, v25.4s, v13.4s\n"
+ "add v26.4s, v26.4s, v13.4s\n"
+ "add v27.4s, v27.4s, v13.4s\n"
+ "add v28.4s, v28.4s, v13.4s\n"
+ "add v29.4s, v29.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v13.4s\n"
+ "add v31.4s, v31.4s, v13.4s\n"
"smin v24.4s, v24.4s, v15.4s\n"
"smin v25.4s, v25.4s, v15.4s\n"
"smin v26.4s, v26.4s, v15.4s\n"
@@ -347,14 +348,14 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
"smin v29.4s, v29.4s, v15.4s\n"
"smin v30.4s, v30.4s, v15.4s\n"
"smin v31.4s, v31.4s, v15.4s\n"
- "smax v24.4s, v24.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v26.4s, v26.4s, v13.4s\n"
- "smax v27.4s, v27.4s, v13.4s\n"
- "smax v28.4s, v28.4s, v13.4s\n"
- "smax v29.4s, v29.4s, v13.4s\n"
- "smax v30.4s, v30.4s, v13.4s\n"
- "smax v31.4s, v31.4s, v13.4s\n"
+ "smax v24.4s, v24.4s, v14.4s\n"
+ "smax v25.4s, v25.4s, v14.4s\n"
+ "smax v26.4s, v26.4s, v14.4s\n"
+ "smax v27.4s, v27.4s, v14.4s\n"
+ "smax v28.4s, v28.4s, v14.4s\n"
+ "smax v29.4s, v29.4s, v14.4s\n"
+ "smax v30.4s, v30.4s, v14.4s\n"
+ "smax v31.4s, v31.4s, v14.4s\n"
"uzp1 v24.16b, v24.16b, v24.16b\n"
"uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v26.16b, v26.16b, v26.16b\n"
@@ -388,14 +389,14 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
"dup v30.4s, v23.s[2]\n"
"dup v31.4s, v23.s[3]\n"
"add x28, x28, #0x4\n"
- "add v24.4s, v24.4s, v12.4s\n"
- "add v25.4s, v25.4s, v12.4s\n"
- "add v26.4s, v26.4s, v12.4s\n"
- "add v27.4s, v27.4s, v12.4s\n"
- "add v28.4s, v28.4s, v12.4s\n"
- "add v29.4s, v29.4s, v12.4s\n"
- "add v30.4s, v30.4s, v12.4s\n"
- "add v31.4s, v31.4s, v12.4s\n"
+ "add v24.4s, v24.4s, v20.4s\n"
+ "add v25.4s, v25.4s, v20.4s\n"
+ "add v26.4s, v26.4s, v20.4s\n"
+ "add v27.4s, v27.4s, v20.4s\n"
+ "add v28.4s, v28.4s, v20.4s\n"
+ "add v29.4s, v29.4s, v20.4s\n"
+ "add v30.4s, v30.4s, v20.4s\n"
+ "add v31.4s, v31.4s, v20.4s\n"
"bgt 1b\n"
"2:" // Tail
"ldr q21, [%x[params], #0x60]\n"
@@ -420,7 +421,7 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
"add x21, x21, x28\n"
".inst 0x6f83e11e // udot v30.4s, v8.16b, v3.4b[0]\n"
".inst 0x6f83e91f // udot v31.4s, v8.16b, v3.4b[2]\n"
- "ldr q8, [%x[params], #0x0]\n"
+ "ldr q17, [%x[params], #0x0]\n"
"add x20, x20, x28\n"
".inst 0x6f81e158 // udot v24.4s, v10.16b, v1.4b[0]\n"
".inst 0x6f81e959 // udot v25.4s, v10.16b, v1.4b[2]\n"
@@ -430,7 +431,7 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
".inst 0x6fa2e93d // udot v29.4s, v9.16b, v2.4b[3]\n"
".inst 0x6fa3e13e // udot v30.4s, v9.16b, v3.4b[1]\n"
".inst 0x6fa3e93f // udot v31.4s, v9.16b, v3.4b[3]\n"
- "ldr q9, [%x[params], #0x10]\n"
+ "ldr q16, [%x[params], #0x10]\n"
".inst 0x6fa1e178 // udot v24.4s, v11.16b, v1.4b[1]\n"
".inst 0x6fa1e979 // udot v25.4s, v11.16b, v1.4b[3]\n"
".inst 0x6fa2e17a // udot v26.4s, v11.16b, v2.4b[1]\n"
@@ -439,68 +440,68 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
".inst 0x6f83e95d // udot v29.4s, v10.16b, v3.4b[2]\n"
".inst 0x6f84e15e // udot v30.4s, v10.16b, v4.4b[0]\n"
".inst 0x6f84e95f // udot v31.4s, v10.16b, v4.4b[2]\n"
- "ldr q10, [%x[params], #0x20]\n"
- ".inst 0x6f82e118 // udot v24.4s, v8.16b, v2.4b[0]\n"
- ".inst 0x6f82e919 // udot v25.4s, v8.16b, v2.4b[2]\n"
- ".inst 0x6f83e11a // udot v26.4s, v8.16b, v3.4b[0]\n"
- ".inst 0x6f83e91b // udot v27.4s, v8.16b, v3.4b[2]\n"
+ "ldr q19, [%x[params], #0x20]\n"
+ ".inst 0x6f82e238 // udot v24.4s, v17.16b, v2.4b[0]\n"
+ ".inst 0x6f82ea39 // udot v25.4s, v17.16b, v2.4b[2]\n"
+ ".inst 0x6f83e23a // udot v26.4s, v17.16b, v3.4b[0]\n"
+ ".inst 0x6f83ea3b // udot v27.4s, v17.16b, v3.4b[2]\n"
".inst 0x6fa3e17c // udot v28.4s, v11.16b, v3.4b[1]\n"
".inst 0x6fa3e97d // udot v29.4s, v11.16b, v3.4b[3]\n"
".inst 0x6fa4e17e // udot v30.4s, v11.16b, v4.4b[1]\n"
".inst 0x6fa4e97f // udot v31.4s, v11.16b, v4.4b[3]\n"
- "ldr q11, [%x[params], #0x30]\n"
- ".inst 0x6fa2e138 // udot v24.4s, v9.16b, v2.4b[1]\n"
- ".inst 0x6fa2e939 // udot v25.4s, v9.16b, v2.4b[3]\n"
- ".inst 0x6fa3e13a // udot v26.4s, v9.16b, v3.4b[1]\n"
- ".inst 0x6fa3e93b // udot v27.4s, v9.16b, v3.4b[3]\n"
- ".inst 0x6f84e11c // udot v28.4s, v8.16b, v4.4b[0]\n"
- ".inst 0x6f84e91d // udot v29.4s, v8.16b, v4.4b[2]\n"
- ".inst 0x6f85e11e // udot v30.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x6f85e91f // udot v31.4s, v8.16b, v5.4b[2]\n"
- "ldr q8, [%x[params], #0x40]\n"
- ".inst 0x6f83e158 // udot v24.4s, v10.16b, v3.4b[0]\n"
- ".inst 0x6f83e959 // udot v25.4s, v10.16b, v3.4b[2]\n"
- ".inst 0x6f84e15a // udot v26.4s, v10.16b, v4.4b[0]\n"
- ".inst 0x6f84e95b // udot v27.4s, v10.16b, v4.4b[2]\n"
- ".inst 0x6fa4e13c // udot v28.4s, v9.16b, v4.4b[1]\n"
- ".inst 0x6fa4e93d // udot v29.4s, v9.16b, v4.4b[3]\n"
- ".inst 0x6fa5e13e // udot v30.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x6fa5e93f // udot v31.4s, v9.16b, v5.4b[3]\n"
- "ldr q9, [%x[params], #0x50]\n"
+ "ldr q18, [%x[params], #0x30]\n"
+ ".inst 0x6fa2e218 // udot v24.4s, v16.16b, v2.4b[1]\n"
+ ".inst 0x6fa2ea19 // udot v25.4s, v16.16b, v2.4b[3]\n"
+ ".inst 0x6fa3e21a // udot v26.4s, v16.16b, v3.4b[1]\n"
+ ".inst 0x6fa3ea1b // udot v27.4s, v16.16b, v3.4b[3]\n"
+ ".inst 0x6f84e23c // udot v28.4s, v17.16b, v4.4b[0]\n"
+ ".inst 0x6f84ea3d // udot v29.4s, v17.16b, v4.4b[2]\n"
+ ".inst 0x6f85e23e // udot v30.4s, v17.16b, v5.4b[0]\n"
+ ".inst 0x6f85ea3f // udot v31.4s, v17.16b, v5.4b[2]\n"
+ "ldr q17, [%x[params], #0x40]\n"
+ ".inst 0x6f83e278 // udot v24.4s, v19.16b, v3.4b[0]\n"
+ ".inst 0x6f83ea79 // udot v25.4s, v19.16b, v3.4b[2]\n"
+ ".inst 0x6f84e27a // udot v26.4s, v19.16b, v4.4b[0]\n"
+ ".inst 0x6f84ea7b // udot v27.4s, v19.16b, v4.4b[2]\n"
+ ".inst 0x6fa4e21c // udot v28.4s, v16.16b, v4.4b[1]\n"
+ ".inst 0x6fa4ea1d // udot v29.4s, v16.16b, v4.4b[3]\n"
+ ".inst 0x6fa5e21e // udot v30.4s, v16.16b, v5.4b[1]\n"
+ ".inst 0x6fa5ea1f // udot v31.4s, v16.16b, v5.4b[3]\n"
+ "ldr q16, [%x[params], #0x50]\n"
"add %x[params], %x[params], #0x80\n"
- ".inst 0x6fa3e178 // udot v24.4s, v11.16b, v3.4b[1]\n"
- ".inst 0x6fa3e979 // udot v25.4s, v11.16b, v3.4b[3]\n"
- ".inst 0x6fa4e17a // udot v26.4s, v11.16b, v4.4b[1]\n"
- ".inst 0x6fa4e97b // udot v27.4s, v11.16b, v4.4b[3]\n"
- ".inst 0x6f85e15c // udot v28.4s, v10.16b, v5.4b[0]\n"
- ".inst 0x6f85e95d // udot v29.4s, v10.16b, v5.4b[2]\n"
- ".inst 0x6f86e15e // udot v30.4s, v10.16b, v6.4b[0]\n"
- ".inst 0x6f86e95f // udot v31.4s, v10.16b, v6.4b[2]\n"
- ".inst 0x6f84e118 // udot v24.4s, v8.16b, v4.4b[0]\n"
- ".inst 0x6f84e919 // udot v25.4s, v8.16b, v4.4b[2]\n"
- ".inst 0x6f85e11a // udot v26.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x6f85e91b // udot v27.4s, v8.16b, v5.4b[2]\n"
- ".inst 0x6fa5e17c // udot v28.4s, v11.16b, v5.4b[1]\n"
- ".inst 0x6fa5e97d // udot v29.4s, v11.16b, v5.4b[3]\n"
- ".inst 0x6fa6e17e // udot v30.4s, v11.16b, v6.4b[1]\n"
- ".inst 0x6fa6e97f // udot v31.4s, v11.16b, v6.4b[3]\n"
- ".inst 0x6fa4e138 // udot v24.4s, v9.16b, v4.4b[1]\n"
- ".inst 0x6fa4e939 // udot v25.4s, v9.16b, v4.4b[3]\n"
+ ".inst 0x6fa3e258 // udot v24.4s, v18.16b, v3.4b[1]\n"
+ ".inst 0x6fa3ea59 // udot v25.4s, v18.16b, v3.4b[3]\n"
+ ".inst 0x6fa4e25a // udot v26.4s, v18.16b, v4.4b[1]\n"
+ ".inst 0x6fa4ea5b // udot v27.4s, v18.16b, v4.4b[3]\n"
+ ".inst 0x6f85e27c // udot v28.4s, v19.16b, v5.4b[0]\n"
+ ".inst 0x6f85ea7d // udot v29.4s, v19.16b, v5.4b[2]\n"
+ ".inst 0x6f86e27e // udot v30.4s, v19.16b, v6.4b[0]\n"
+ ".inst 0x6f86ea7f // udot v31.4s, v19.16b, v6.4b[2]\n"
+ ".inst 0x6f84e238 // udot v24.4s, v17.16b, v4.4b[0]\n"
+ ".inst 0x6f84ea39 // udot v25.4s, v17.16b, v4.4b[2]\n"
+ ".inst 0x6f85e23a // udot v26.4s, v17.16b, v5.4b[0]\n"
+ ".inst 0x6f85ea3b // udot v27.4s, v17.16b, v5.4b[2]\n"
+ ".inst 0x6fa5e25c // udot v28.4s, v18.16b, v5.4b[1]\n"
+ ".inst 0x6fa5ea5d // udot v29.4s, v18.16b, v5.4b[3]\n"
+ ".inst 0x6fa6e25e // udot v30.4s, v18.16b, v6.4b[1]\n"
+ ".inst 0x6fa6ea5f // udot v31.4s, v18.16b, v6.4b[3]\n"
+ ".inst 0x6fa4e218 // udot v24.4s, v16.16b, v4.4b[1]\n"
+ ".inst 0x6fa4ea19 // udot v25.4s, v16.16b, v4.4b[3]\n"
"sqrdmulh v24.4s, v24.4s, v21.4s\n"
- ".inst 0x6fa5e13a // udot v26.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x6fa5e93b // udot v27.4s, v9.16b, v5.4b[3]\n"
+ ".inst 0x6fa5e21a // udot v26.4s, v16.16b, v5.4b[1]\n"
+ ".inst 0x6fa5ea1b // udot v27.4s, v16.16b, v5.4b[3]\n"
"sqrdmulh v25.4s, v25.4s, v21.4s\n"
- ".inst 0x6f86e11c // udot v28.4s, v8.16b, v6.4b[0]\n"
- ".inst 0x6f86e91d // udot v29.4s, v8.16b, v6.4b[2]\n"
+ ".inst 0x6f86e23c // udot v28.4s, v17.16b, v6.4b[0]\n"
+ ".inst 0x6f86ea3d // udot v29.4s, v17.16b, v6.4b[2]\n"
"sqrdmulh v26.4s, v26.4s, v21.4s\n"
- ".inst 0x6f87e11e // udot v30.4s, v8.16b, v7.4b[0]\n"
- ".inst 0x6f87e91f // udot v31.4s, v8.16b, v7.4b[2]\n"
+ ".inst 0x6f87e23e // udot v30.4s, v17.16b, v7.4b[0]\n"
+ ".inst 0x6f87ea3f // udot v31.4s, v17.16b, v7.4b[2]\n"
"sqrdmulh v27.4s, v27.4s, v21.4s\n"
- ".inst 0x6fa6e13c // udot v28.4s, v9.16b, v6.4b[1]\n"
- ".inst 0x6fa6e93d // udot v29.4s, v9.16b, v6.4b[3]\n"
+ ".inst 0x6fa6e21c // udot v28.4s, v16.16b, v6.4b[1]\n"
+ ".inst 0x6fa6ea1d // udot v29.4s, v16.16b, v6.4b[3]\n"
"and v19.16b, v24.16b, v20.16b\n"
- ".inst 0x6fa7e13e // udot v30.4s, v9.16b, v7.4b[1]\n"
- ".inst 0x6fa7e93f // udot v31.4s, v9.16b, v7.4b[3]\n"
+ ".inst 0x6fa7e21e // udot v30.4s, v16.16b, v7.4b[1]\n"
+ ".inst 0x6fa7ea1f // udot v31.4s, v16.16b, v7.4b[3]\n"
"and v18.16b, v25.16b, v20.16b\n"
"and v17.16b, v26.16b, v20.16b\n"
"and v16.16b, v27.16b, v20.16b\n"
@@ -536,14 +537,14 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
"srshl v29.4s, v29.4s, v20.4s\n"
"srshl v30.4s, v30.4s, v20.4s\n"
"srshl v31.4s, v31.4s, v20.4s\n"
- "add v24.4s, v24.4s, v14.4s\n"
- "add v25.4s, v25.4s, v14.4s\n"
- "add v26.4s, v26.4s, v14.4s\n"
- "add v27.4s, v27.4s, v14.4s\n"
- "add v28.4s, v28.4s, v14.4s\n"
- "add v29.4s, v29.4s, v14.4s\n"
- "add v30.4s, v30.4s, v14.4s\n"
- "add v31.4s, v31.4s, v14.4s\n"
+ "add v24.4s, v24.4s, v13.4s\n"
+ "add v25.4s, v25.4s, v13.4s\n"
+ "add v26.4s, v26.4s, v13.4s\n"
+ "add v27.4s, v27.4s, v13.4s\n"
+ "add v28.4s, v28.4s, v13.4s\n"
+ "add v29.4s, v29.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v13.4s\n"
+ "add v31.4s, v31.4s, v13.4s\n"
"smin v24.4s, v24.4s, v15.4s\n"
"smin v25.4s, v25.4s, v15.4s\n"
"smin v26.4s, v26.4s, v15.4s\n"
@@ -552,14 +553,14 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
"smin v29.4s, v29.4s, v15.4s\n"
"smin v30.4s, v30.4s, v15.4s\n"
"smin v31.4s, v31.4s, v15.4s\n"
- "smax v24.4s, v24.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v26.4s, v26.4s, v13.4s\n"
- "smax v27.4s, v27.4s, v13.4s\n"
- "smax v28.4s, v28.4s, v13.4s\n"
- "smax v29.4s, v29.4s, v13.4s\n"
- "smax v30.4s, v30.4s, v13.4s\n"
- "smax v31.4s, v31.4s, v13.4s\n"
+ "smax v24.4s, v24.4s, v14.4s\n"
+ "smax v25.4s, v25.4s, v14.4s\n"
+ "smax v26.4s, v26.4s, v14.4s\n"
+ "smax v27.4s, v27.4s, v14.4s\n"
+ "smax v28.4s, v28.4s, v14.4s\n"
+ "smax v29.4s, v29.4s, v14.4s\n"
+ "smax v30.4s, v30.4s, v14.4s\n"
+ "smax v31.4s, v31.4s, v14.4s\n"
"uzp1 v24.16b, v24.16b, v24.16b\n"
"uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v26.16b, v26.16b, v26.16b\n"
@@ -635,4 +636,5 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
} // namespace depthwise
} // namespace arm_conv
-#endif // defined(__aarch64__)
+
+#endif // defined(__aarch64__)