aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp2738
1 files changed, 1365 insertions, 1373 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
index 986937f3b4..5a28daffbf 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
@@ -30,15 +30,7 @@
namespace arm_conv {
namespace depthwise {
-void a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst_impl(
- const unsigned int n_channels,
- const uint8_t *const *const inptrs,
- const uint8_t *params,
- const int32_t *, // Bias, should be wrapped into the parameters
- const arm_gemm::Requantize32& qp,
- const int32_t *, const int32_t *, // Requant parameters, also wrapped
- uint8_t *const *const outptrs
-)
+void a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst_impl(const unsigned int n_channels, const uint8_t *const *const inptrs, const uint8_t *params, const int32_t *, const arm_gemm::Requantize32& qp, const int32_t *, const int32_t *, uint8_t *const *const outptrs)
{
__asm__ __volatile__(
"mov x20, #0x1\n"
@@ -47,817 +39,817 @@ void a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst_impl(
"ldp x13, x12, [%x[inptrs], #0x10]\n"
"orr x20, x20, #0x10000\n"
"lsr x11, %x[n_channels], #0x4\n"
- "dup v14.4s, w20\n"
+ "dup v12.4s, w20\n"
"ldp x10, x9, [%x[inptrs], #0x20]\n"
"add x20, %x[qp], %[offsetof_Requantize32_minval]\n"
"ld1r { v13.4s }, [x20]\n"
"add x20, %x[qp], %[offsetof_Requantize32_maxval]\n"
- "ld1r { v12.4s }, [x20]\n"
- "add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n"
"ld1r { v11.4s }, [x20]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n"
+ "ld1r { v16.4s }, [x20]\n"
"add x20, %x[qp], %[offsetof_Requantize32_c_offset]\n"
- "ld1r { v10.4s }, [x20]\n"
+ "ld1r { v14.4s }, [x20]\n"
"mov x28, #0x0\n"
"mov x27, #0x0\n"
- "ldp x26, x25, [%x[inptrs], #0x30]\n"
- "ldp x24, x23, [%x[outptrs], #0x0]\n"
- "ldp x22, x21, [%x[outptrs], #0x10]\n"
+ "ldp x26, x21, [%x[inptrs], #0x30]\n"
+ "ldp x25, x24, [%x[outptrs], #0x0]\n"
+ "ldp x23, x22, [%x[outptrs], #0x10]\n"
"cbz x11, 3f\n"
- "ldr q9, [x15, x28]\n"
- "ldr q8, [x14, x28]\n"
- "subs x11, x11, #0x1\n"
- "ldr q7, [x13, x28]\n"
- "ldr q6, [x12, x28]\n"
- "zip2 v5.16b, v9.16b, v7.16b\n"
- "zip1 v9.16b, v9.16b, v7.16b\n"
- "ldr q4, [x10, x28]\n"
- "ldr q3, [x9, x28]\n"
- "zip1 v7.16b, v8.16b, v6.16b\n"
- "zip2 v6.16b, v8.16b, v6.16b\n"
- "ldr q2, [x26, x28]\n"
- "ldr q1, [x25, x28]\n"
- "zip2 v8.16b, v9.16b, v7.16b\n"
- "zip1 v9.16b, v9.16b, v7.16b\n"
- "ldr q0, [%x[params], #0x10]\n"
- "ldr q16, [%x[params], #0x20]\n"
- "zip1 v7.16b, v5.16b, v6.16b\n"
- "zip2 v6.16b, v5.16b, v6.16b\n"
- "ldr q5, [%x[params], #0x0]\n"
- "ldr q31, [%x[params], #0x30]\n"
- "zip2 v30.16b, v4.16b, v2.16b\n"
- "zip1 v4.16b, v4.16b, v2.16b\n"
- "ldp x15, x14, [%x[inptrs], #0x40]\n"
- "ldr q29, [x15, x28]\n"
- "zip1 v2.16b, v3.16b, v1.16b\n"
- "zip2 v1.16b, v3.16b, v1.16b\n"
+ "ldr q15, [x15, x28]\n"
"ldr q28, [x14, x28]\n"
- "ldp x13, x12, [%x[inptrs], #0x50]\n"
- "zip2 v3.16b, v4.16b, v2.16b\n"
- "zip1 v4.16b, v4.16b, v2.16b\n"
- "ldr q27, [x13, x28]\n"
- "ldr q26, [x12, x28]\n"
- "zip2 v25.16b, v29.16b, v27.16b\n"
- "zip1 v29.16b, v29.16b, v27.16b\n"
- "ldp x10, x9, [%x[inptrs], #0x60]\n"
- "ldr q24, [x10, x28]\n"
- "zip1 v27.16b, v28.16b, v26.16b\n"
- "zip2 v26.16b, v28.16b, v26.16b\n"
- "ldr q23, [x9, x28]\n"
- "ldp x26, x25, [%x[inptrs], #0x70]\n"
- "zip1 v2.16b, v30.16b, v1.16b\n"
- "zip2 v1.16b, v30.16b, v1.16b\n"
- "ldr q22, [x26, x28]\n"
- "ldr q21, [x25, x28]\n"
- "zip2 v20.16b, v24.16b, v22.16b\n"
- "zip1 v24.16b, v24.16b, v22.16b\n"
- "zip1 v22.16b, v23.16b, v21.16b\n"
- "zip2 v21.16b, v23.16b, v21.16b\n"
+ "subs x11, x11, #0x1\n"
+ "ldr q30, [x13, x28]\n"
+ "ldr q8, [x12, x28]\n"
+ "zip2 v19.16b, v15.16b, v30.16b\n"
+ "zip1 v15.16b, v15.16b, v30.16b\n"
+ "ldr q26, [x10, x28]\n"
+ "ldr q0, [x9, x28]\n"
+ "zip1 v7.16b, v28.16b, v8.16b\n"
+ "zip2 v8.16b, v28.16b, v8.16b\n"
+ "ldr q29, [x26, x28]\n"
+ "ldr q10, [x21, x28]\n"
+ "zip2 v25.16b, v15.16b, v7.16b\n"
+ "zip1 v15.16b, v15.16b, v7.16b\n"
+ "ldr q1, [%x[params], #0x10]\n"
+ "ldr q6, [%x[params], #0x20]\n"
+ "zip1 v7.16b, v19.16b, v8.16b\n"
+ "zip2 v8.16b, v19.16b, v8.16b\n"
+ "ldr q31, [%x[params], #0x0]\n"
+ "ldr q20, [%x[params], #0x30]\n"
+ "zip2 v21.16b, v26.16b, v29.16b\n"
+ "zip1 v26.16b, v26.16b, v29.16b\n"
+ "ldp x21, x20, [%x[inptrs], #0x40]\n"
+ "ldr q22, [x21, x28]\n"
+ "zip1 v27.16b, v0.16b, v10.16b\n"
+ "zip2 v10.16b, v0.16b, v10.16b\n"
+ "ldr q17, [x20, x28]\n"
+ "ldp x21, x20, [%x[inptrs], #0x50]\n"
+ "zip2 v23.16b, v26.16b, v27.16b\n"
+ "zip1 v26.16b, v26.16b, v27.16b\n"
+ "ldr q9, [x21, x28]\n"
+ "ldr q5, [x20, x28]\n"
+ "zip2 v28.16b, v22.16b, v9.16b\n"
+ "zip1 v22.16b, v22.16b, v9.16b\n"
+ "ldp x21, x20, [%x[inptrs], #0x60]\n"
+ "ldr q27, [x21, x28]\n"
+ "zip1 v24.16b, v17.16b, v5.16b\n"
+ "zip2 v5.16b, v17.16b, v5.16b\n"
+ "ldr q18, [x20, x28]\n"
+ "ldp x21, x20, [%x[inptrs], #0x70]\n"
+ "zip1 v3.16b, v21.16b, v10.16b\n"
+ "zip2 v10.16b, v21.16b, v10.16b\n"
+ "ldr q4, [x21, x28]\n"
+ "ldr q9, [x20, x28]\n"
+ "zip2 v17.16b, v27.16b, v4.16b\n"
+ "zip1 v27.16b, v27.16b, v4.16b\n"
+ "zip1 v4.16b, v18.16b, v9.16b\n"
+ "zip2 v9.16b, v18.16b, v9.16b\n"
"ldp x15, x14, [%x[inptrs], #0x0]\n"
"ldp x13, x12, [%x[inptrs], #0x10]\n"
"ldp x10, x9, [%x[inptrs], #0x20]\n"
- "ldp x26, x25, [%x[inptrs], #0x30]\n"
- "zip2 v28.16b, v29.16b, v27.16b\n"
- "zip1 v29.16b, v29.16b, v27.16b\n"
- "zip1 v27.16b, v25.16b, v26.16b\n"
- "zip2 v26.16b, v25.16b, v26.16b\n"
+ "ldp x26, x21, [%x[inptrs], #0x30]\n"
+ "zip2 v19.16b, v22.16b, v24.16b\n"
+ "zip1 v22.16b, v22.16b, v24.16b\n"
+ "zip1 v0.16b, v28.16b, v5.16b\n"
+ "zip2 v5.16b, v28.16b, v5.16b\n"
"add %x[params], %x[params], #0x40\n"
- "zip2 v23.16b, v24.16b, v22.16b\n"
- "zip1 v24.16b, v24.16b, v22.16b\n"
- "zip1 v22.16b, v20.16b, v21.16b\n"
- "zip2 v21.16b, v20.16b, v21.16b\n"
- "mov v30.16b, v5.16b\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
+ "zip2 v24.16b, v27.16b, v4.16b\n"
+ "zip1 v27.16b, v27.16b, v4.16b\n"
+ "zip1 v2.16b, v17.16b, v9.16b\n"
+ "zip2 v9.16b, v17.16b, v9.16b\n"
+ "mov v30.16b, v31.16b\n"
+ "mov v29.16b, v31.16b\n"
+ "mov v28.16b, v31.16b\n"
"beq 2f\n"
"1:" // Loop
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8495d3 // udot v19.4s, v14.16b, v4.16b\n"
- ".inst 0x6e899405 // udot v5.4s, v0.16b, v9.16b\n"
+ "movi v21.4s, #0x0\n"
+ ".inst 0x6e9a9595 // udot v21.4s, v12.16b, v26.16b\n"
+ ".inst 0x6e8f943f // udot v31.4s, v1.16b, v15.16b\n"
"add x28, x28, #0x10\n"
- ".inst 0x6e9d95d3 // udot v19.4s, v14.16b, v29.16b\n"
- ".inst 0x6e849419 // udot v25.4s, v0.16b, v4.16b\n"
+ ".inst 0x6e969595 // udot v21.4s, v12.16b, v22.16b\n"
+ ".inst 0x6e9a943d // udot v29.4s, v1.16b, v26.16b\n"
+ "movi v18.4s, #0x0\n"
"subs x11, x11, #0x1\n"
- ".inst 0x6e849605 // udot v5.4s, v16.16b, v4.16b\n"
- "ext v4.16b, v4.16b, v4.16b, #0x1\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9895d2 // udot v18.4s, v14.16b, v24.16b\n"
- ".inst 0x6e8995d3 // udot v19.4s, v14.16b, v9.16b\n"
- "ext v9.16b, v9.16b, v9.16b, #0x1\n"
- ".inst 0x6e9d9619 // udot v25.4s, v16.16b, v29.16b\n"
- ".inst 0x6e9d97e5 // udot v5.4s, v31.16b, v29.16b\n"
- "ext v29.16b, v29.16b, v29.16b, #0x1\n"
- ".inst 0x6e89941e // udot v30.4s, v0.16b, v9.16b\n"
- ".inst 0x6e849414 // udot v20.4s, v0.16b, v4.16b\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x6e8495d1 // udot v17.4s, v14.16b, v4.16b\n"
- ".inst 0x6e9d95d1 // udot v17.4s, v14.16b, v29.16b\n"
- ".inst 0x6e9897f9 // udot v25.4s, v31.16b, v24.16b\n"
- "ext v24.16b, v24.16b, v24.16b, #0x1\n"
- ".inst 0x6e84961e // udot v30.4s, v16.16b, v4.16b\n"
- "ldr q4, [%x[params], #0x10]\n"
- ".inst 0x6e9d9614 // udot v20.4s, v16.16b, v29.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9895d0 // udot v16.4s, v14.16b, v24.16b\n"
- ".inst 0x6e8995d1 // udot v17.4s, v14.16b, v9.16b\n"
- "ldr q9, [%x[params], #0x0]\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- ".inst 0x6e9d97fe // udot v30.4s, v31.16b, v29.16b\n"
- ".inst 0x6e9897f4 // udot v20.4s, v31.16b, v24.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "ldr q9, [%x[params], #0x60]\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "ldr q16, [%x[params], #0x40]\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "ldr q31, [%x[params], #0x50]\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "ldr q0, [%x[params], #0x30]\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "ldr q4, [%x[params], #0x70]\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
+ ".inst 0x6e9a94df // udot v31.4s, v6.16b, v26.16b\n"
+ "ext v26.16b, v26.16b, v26.16b, #0x1\n"
+ "mov v17.16b, v21.16b\n .inst 0x6e9b9591 // udot v17.4s, v12.16b, v27.16b\n"
+ ".inst 0x6e8f9595 // udot v21.4s, v12.16b, v15.16b\n"
+ "ext v15.16b, v15.16b, v15.16b, #0x1\n"
+ ".inst 0x6e9a9592 // udot v18.4s, v12.16b, v26.16b\n"
+ ".inst 0x6e9694dd // udot v29.4s, v6.16b, v22.16b\n"
+ ".inst 0x6e96969f // udot v31.4s, v20.16b, v22.16b\n"
+ "ext v22.16b, v22.16b, v22.16b, #0x1\n"
+ ".inst 0x6e8f943e // udot v30.4s, v1.16b, v15.16b\n"
+ ".inst 0x6e9a943c // udot v28.4s, v1.16b, v26.16b\n"
+ "mls v31.4s, v21.4s, v16.4s\n"
+ ".inst 0x6e969592 // udot v18.4s, v12.16b, v22.16b\n"
+ ".inst 0x6e9b969d // udot v29.4s, v20.16b, v27.16b\n"
+ "ext v27.16b, v27.16b, v27.16b, #0x1\n"
+ ".inst 0x6e9a94de // udot v30.4s, v6.16b, v26.16b\n"
+ "ldr q26, [%x[params], #0x10]\n"
+ ".inst 0x6e9694dc // udot v28.4s, v6.16b, v22.16b\n"
+ "mls v29.4s, v17.4s, v16.4s\n"
+ "mov v21.16b, v18.16b\n .inst 0x6e9b9595 // udot v21.4s, v12.16b, v27.16b\n"
+ ".inst 0x6e8f9592 // udot v18.4s, v12.16b, v15.16b\n"
+ "ldr q17, [%x[params], #0x0]\n"
+ "sqrdmulh v31.4s, v31.4s, v17.4s\n"
+ ".inst 0x6e96969e // udot v30.4s, v20.16b, v22.16b\n"
+ ".inst 0x6e9b969c // udot v28.4s, v20.16b, v27.16b\n"
+ "mls v30.4s, v18.4s, v16.4s\n"
+ "mls v28.4s, v21.4s, v16.4s\n"
+ "and v15.16b, v31.16b, v26.16b\n"
+ "sshr v15.4s, v15.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v17.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v17.4s\n"
+ "sqrdmulh v28.4s, v28.4s, v17.4s\n"
+ "ldr q1, [%x[params], #0x60]\n"
+ "sqadd v31.4s, v31.4s, v15.4s\n"
+ "and v18.16b, v30.16b, v26.16b\n"
+ "and v21.16b, v29.16b, v26.16b\n"
+ "and v17.16b, v28.16b, v26.16b\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "srshl v31.4s, v31.4s, v26.4s\n"
+ "sqadd v30.4s, v30.4s, v18.4s\n"
+ "ldr q18, [%x[params], #0x40]\n"
+ "sqadd v29.4s, v29.4s, v21.4s\n"
+ "ldr q27, [%x[params], #0x50]\n"
+ "sqadd v28.4s, v28.4s, v17.4s\n"
+ "ldr q15, [%x[params], #0x30]\n"
+ "add v31.4s, v31.4s, v14.4s\n"
+ "srshl v30.4s, v30.4s, v26.4s\n"
+ "srshl v29.4s, v29.4s, v26.4s\n"
+ "srshl v28.4s, v28.4s, v26.4s\n"
+ "ldr q20, [%x[params], #0x70]\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v14.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v28.4s, v28.4s, v14.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
"smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8395d3 // udot v19.4s, v14.16b, v3.16b\n"
- ".inst 0x6e9c95d3 // udot v19.4s, v14.16b, v28.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smin v30.4s, v30.4s, v11.4s\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v28.4s, v28.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
+ "movi v22.4s, #0x0\n"
+ ".inst 0x6e979596 // udot v22.4s, v12.16b, v23.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s5, [x24, x27]\n"
- "ldr q5, [%x[params], #0x20]\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9795d2 // udot v18.4s, v14.16b, v23.16b\n"
+ "str s31, [x25, x27]\n"
+ "ldr q26, [%x[params], #0x20]\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ ".inst 0x6e939596 // udot v22.4s, v12.16b, v19.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
- ".inst 0x6e8895d3 // udot v19.4s, v14.16b, v8.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "mov v30.16b, v5.16b\n"
- "str s20, [x21, x27]\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e889405 // udot v5.4s, v0.16b, v8.16b\n"
- ".inst 0x6e839419 // udot v25.4s, v0.16b, v3.16b\n"
- ".inst 0x6e839605 // udot v5.4s, v16.16b, v3.16b\n"
- "ext v8.16b, v8.16b, v8.16b, #0x1\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "str s30, [x24, x27]\n"
+ "mov v6.16b, v22.16b\n .inst 0x6e989586 // udot v6.4s, v12.16b, v24.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "str s29, [x23, x27]\n"
+ "mov v30.16b, v26.16b\n"
+ ".inst 0x6e999596 // udot v22.4s, v12.16b, v25.16b\n"
+ "str s28, [x22, x27]\n"
+ "mov v29.16b, v26.16b\n"
+ "mov v21.16b, v26.16b\n"
+ ".inst 0x6e9995fa // udot v26.4s, v15.16b, v25.16b\n"
+ ".inst 0x6e9795fd // udot v29.4s, v15.16b, v23.16b\n"
+ ".inst 0x6e97965a // udot v26.4s, v18.16b, v23.16b\n"
+ "ext v25.16b, v25.16b, v25.16b, #0x1\n"
"add x27, x27, #0x4\n"
- "ext v3.16b, v3.16b, v3.16b, #0x1\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x6e88941e // udot v30.4s, v0.16b, v8.16b\n"
- ".inst 0x6e839414 // udot v20.4s, v0.16b, v3.16b\n"
- ".inst 0x6e8395d1 // udot v17.4s, v14.16b, v3.16b\n"
- ".inst 0x6e9c9619 // udot v25.4s, v16.16b, v28.16b\n"
- ".inst 0x6e9c97e5 // udot v5.4s, v31.16b, v28.16b\n"
- "ext v28.16b, v28.16b, v28.16b, #0x1\n"
- ".inst 0x6e83961e // udot v30.4s, v16.16b, v3.16b\n"
- "ldr q3, [x9, x28]\n"
- ".inst 0x6e9c9614 // udot v20.4s, v16.16b, v28.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9c95d1 // udot v17.4s, v14.16b, v28.16b\n"
- ".inst 0x6e9797f9 // udot v25.4s, v31.16b, v23.16b\n"
"ext v23.16b, v23.16b, v23.16b, #0x1\n"
- ".inst 0x6e9c97fe // udot v30.4s, v31.16b, v28.16b\n"
- ".inst 0x6e9797f4 // udot v20.4s, v31.16b, v23.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9795d0 // udot v16.4s, v14.16b, v23.16b\n"
- ".inst 0x6e8895d1 // udot v17.4s, v14.16b, v8.16b\n"
- "ldr q8, [x14, x28]\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "ldr q9, [%x[params], #0xc0]\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "ldr q16, [%x[params], #0xa0]\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "ldr q31, [%x[params], #0xb0]\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "ldr q0, [%x[params], #0x90]\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "ldr q4, [%x[params], #0xd0]\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
+ "movi v28.4s, #0x0\n"
+ ".inst 0x6e9995fe // udot v30.4s, v15.16b, v25.16b\n"
+ ".inst 0x6e9795f5 // udot v21.4s, v15.16b, v23.16b\n"
+ ".inst 0x6e97959c // udot v28.4s, v12.16b, v23.16b\n"
+ ".inst 0x6e93965d // udot v29.4s, v18.16b, v19.16b\n"
+ ".inst 0x6e93977a // udot v26.4s, v27.16b, v19.16b\n"
+ "ext v19.16b, v19.16b, v19.16b, #0x1\n"
+ ".inst 0x6e97965e // udot v30.4s, v18.16b, v23.16b\n"
+ "ldr q4, [x9, x28]\n"
+ ".inst 0x6e939655 // udot v21.4s, v18.16b, v19.16b\n"
+ "mls v26.4s, v22.4s, v16.4s\n"
+ ".inst 0x6e93959c // udot v28.4s, v12.16b, v19.16b\n"
+ ".inst 0x6e98977d // udot v29.4s, v27.16b, v24.16b\n"
+ "ext v24.16b, v24.16b, v24.16b, #0x1\n"
+ ".inst 0x6e93977e // udot v30.4s, v27.16b, v19.16b\n"
+ ".inst 0x6e989775 // udot v21.4s, v27.16b, v24.16b\n"
+ "sqrdmulh v26.4s, v26.4s, v1.4s\n"
+ "mov v17.16b, v28.16b\n .inst 0x6e989591 // udot v17.4s, v12.16b, v24.16b\n"
+ ".inst 0x6e99959c // udot v28.4s, v12.16b, v25.16b\n"
+ "ldr q31, [x14, x28]\n"
+ "mls v30.4s, v28.4s, v16.4s\n"
+ "mls v29.4s, v6.4s, v16.4s\n"
+ "mls v21.4s, v17.4s, v16.4s\n"
+ "and v17.16b, v26.16b, v20.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v1.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v1.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v1.4s\n"
+ "ldr q27, [%x[params], #0xc0]\n"
+ "sqadd v26.4s, v26.4s, v17.4s\n"
+ "and v18.16b, v30.16b, v20.16b\n"
+ "and v6.16b, v29.16b, v20.16b\n"
+ "and v17.16b, v21.16b, v20.16b\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v6.4s, v6.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "srshl v26.4s, v26.4s, v20.4s\n"
+ "sqadd v30.4s, v30.4s, v18.4s\n"
+ "ldr q28, [%x[params], #0xa0]\n"
+ "sqadd v29.4s, v29.4s, v6.4s\n"
+ "ldr q24, [%x[params], #0xb0]\n"
+ "sqadd v21.4s, v21.4s, v17.4s\n"
+ "ldr q15, [%x[params], #0x90]\n"
+ "add v26.4s, v26.4s, v14.4s\n"
+ "srshl v30.4s, v30.4s, v20.4s\n"
+ "srshl v29.4s, v29.4s, v20.4s\n"
+ "srshl v21.4s, v21.4s, v20.4s\n"
+ "ldr q1, [%x[params], #0xd0]\n"
+ "smax v26.4s, v26.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v14.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v21.4s, v21.4s, v14.4s\n"
+ "smin v26.4s, v26.4s, v11.4s\n"
"smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8295d3 // udot v19.4s, v14.16b, v2.16b\n"
- ".inst 0x6e9b95d3 // udot v19.4s, v14.16b, v27.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v21.4s, v21.4s, v13.4s\n"
+ "smin v30.4s, v30.4s, v11.4s\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v21.4s, v21.4s, v11.4s\n"
+ "uzp1 v26.16b, v26.16b, v26.16b\n"
+ "movi v22.4s, #0x0\n"
+ ".inst 0x6e839596 // udot v22.4s, v12.16b, v3.16b\n"
+ ".inst 0x6e809596 // udot v22.4s, v12.16b, v0.16b\n"
+ "uzp1 v26.16b, v26.16b, v26.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s5, [x24, x27]\n"
- "ldr q5, [%x[params], #0x80]\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9695d2 // udot v18.4s, v14.16b, v22.16b\n"
+ "str s26, [x25, x27]\n"
+ "ldr q26, [%x[params], #0x80]\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
+ "mov v18.16b, v22.16b\n .inst 0x6e829592 // udot v18.4s, v12.16b, v2.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
- ".inst 0x6e8795d3 // udot v19.4s, v14.16b, v7.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "mov v30.16b, v5.16b\n"
- "str s20, [x21, x27]\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e879405 // udot v5.4s, v0.16b, v7.16b\n"
- ".inst 0x6e829419 // udot v25.4s, v0.16b, v2.16b\n"
- ".inst 0x6e829605 // udot v5.4s, v16.16b, v2.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "str s30, [x24, x27]\n"
+ ".inst 0x6e879596 // udot v22.4s, v12.16b, v7.16b\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
+ "str s29, [x23, x27]\n"
+ "mov v6.16b, v26.16b\n"
+ "str s21, [x22, x27]\n"
+ "mov v25.16b, v26.16b\n"
+ "mov v20.16b, v26.16b\n"
+ ".inst 0x6e8795fa // udot v26.4s, v15.16b, v7.16b\n"
+ ".inst 0x6e8395f9 // udot v25.4s, v15.16b, v3.16b\n"
+ ".inst 0x6e83979a // udot v26.4s, v28.16b, v3.16b\n"
"ext v7.16b, v7.16b, v7.16b, #0x1\n"
"add x27, x27, #0x4\n"
+ "ext v3.16b, v3.16b, v3.16b, #0x1\n"
+ "movi v23.4s, #0x0\n"
+ ".inst 0x6e8795e6 // udot v6.4s, v15.16b, v7.16b\n"
+ ".inst 0x6e8395f4 // udot v20.4s, v15.16b, v3.16b\n"
+ ".inst 0x6e839597 // udot v23.4s, v12.16b, v3.16b\n"
+ ".inst 0x6e809799 // udot v25.4s, v28.16b, v0.16b\n"
+ ".inst 0x6e80971a // udot v26.4s, v24.16b, v0.16b\n"
+ "ext v0.16b, v0.16b, v0.16b, #0x1\n"
+ ".inst 0x6e839786 // udot v6.4s, v28.16b, v3.16b\n"
+ "ldr q19, [x26, x28]\n"
+ ".inst 0x6e809794 // udot v20.4s, v28.16b, v0.16b\n"
+ "mls v26.4s, v22.4s, v16.4s\n"
+ ".inst 0x6e809597 // udot v23.4s, v12.16b, v0.16b\n"
+ ".inst 0x6e829719 // udot v25.4s, v24.16b, v2.16b\n"
"ext v2.16b, v2.16b, v2.16b, #0x1\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x6e87941e // udot v30.4s, v0.16b, v7.16b\n"
- ".inst 0x6e829414 // udot v20.4s, v0.16b, v2.16b\n"
- ".inst 0x6e8295d1 // udot v17.4s, v14.16b, v2.16b\n"
- ".inst 0x6e9b9619 // udot v25.4s, v16.16b, v27.16b\n"
- ".inst 0x6e9b97e5 // udot v5.4s, v31.16b, v27.16b\n"
- "ext v27.16b, v27.16b, v27.16b, #0x1\n"
- ".inst 0x6e82961e // udot v30.4s, v16.16b, v2.16b\n"
- "ldr q2, [x26, x28]\n"
- ".inst 0x6e9b9614 // udot v20.4s, v16.16b, v27.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9b95d1 // udot v17.4s, v14.16b, v27.16b\n"
- ".inst 0x6e9697f9 // udot v25.4s, v31.16b, v22.16b\n"
- "ext v22.16b, v22.16b, v22.16b, #0x1\n"
- ".inst 0x6e9b97fe // udot v30.4s, v31.16b, v27.16b\n"
- ".inst 0x6e9697f4 // udot v20.4s, v31.16b, v22.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9695d0 // udot v16.4s, v14.16b, v22.16b\n"
- ".inst 0x6e8795d1 // udot v17.4s, v14.16b, v7.16b\n"
- "ldr q7, [x13, x28]\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "ldr q9, [%x[params], #0x120]\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "ldr q16, [%x[params], #0x100]\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "ldr q31, [%x[params], #0x110]\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "ldr q0, [%x[params], #0xf0]\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "ldr q4, [%x[params], #0x130]\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smax v30.4s, v30.4s, v13.4s\n"
+ ".inst 0x6e809706 // udot v6.4s, v24.16b, v0.16b\n"
+ ".inst 0x6e829714 // udot v20.4s, v24.16b, v2.16b\n"
+ "sqrdmulh v26.4s, v26.4s, v27.4s\n"
+ "mov v17.16b, v23.16b\n .inst 0x6e829591 // udot v17.4s, v12.16b, v2.16b\n"
+ ".inst 0x6e879597 // udot v23.4s, v12.16b, v7.16b\n"
+ "ldr q21, [x13, x28]\n"
+ "mls v6.4s, v23.4s, v16.4s\n"
+ "mls v25.4s, v18.4s, v16.4s\n"
+ "mls v20.4s, v17.4s, v16.4s\n"
+ "and v17.16b, v26.16b, v1.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v6.4s, v6.4s, v27.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v27.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v27.4s\n"
+ "ldr q15, [%x[params], #0x120]\n"
+ "sqadd v26.4s, v26.4s, v17.4s\n"
+ "and v18.16b, v6.16b, v1.16b\n"
+ "and v22.16b, v25.16b, v1.16b\n"
+ "and v17.16b, v20.16b, v1.16b\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "srshl v26.4s, v26.4s, v1.4s\n"
+ "sqadd v6.4s, v6.4s, v18.4s\n"
+ "ldr q30, [%x[params], #0x100]\n"
+ "sqadd v25.4s, v25.4s, v22.4s\n"
+ "ldr q27, [%x[params], #0x110]\n"
+ "sqadd v20.4s, v20.4s, v17.4s\n"
+ "ldr q24, [%x[params], #0xf0]\n"
+ "add v26.4s, v26.4s, v14.4s\n"
+ "srshl v6.4s, v6.4s, v1.4s\n"
+ "srshl v25.4s, v25.4s, v1.4s\n"
+ "srshl v20.4s, v20.4s, v1.4s\n"
+ "ldr q23, [%x[params], #0x130]\n"
+ "smax v26.4s, v26.4s, v13.4s\n"
+ "add v6.4s, v6.4s, v14.4s\n"
+ "add v25.4s, v25.4s, v14.4s\n"
+ "add v20.4s, v20.4s, v14.4s\n"
+ "smin v26.4s, v26.4s, v11.4s\n"
+ "smax v6.4s, v6.4s, v13.4s\n"
"smax v25.4s, v25.4s, v13.4s\n"
"smax v20.4s, v20.4s, v13.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8195d3 // udot v19.4s, v14.16b, v1.16b\n"
- ".inst 0x6e9a95d3 // udot v19.4s, v14.16b, v26.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s5, [x24, x27]\n"
- "ldr q5, [%x[params], #0xe0]\n"
+ "smin v6.4s, v6.4s, v11.4s\n"
+ "smin v25.4s, v25.4s, v11.4s\n"
+ "smin v20.4s, v20.4s, v11.4s\n"
+ "uzp1 v26.16b, v26.16b, v26.16b\n"
+ "movi v0.4s, #0x0\n"
+ ".inst 0x6e8a9580 // udot v0.4s, v12.16b, v10.16b\n"
+ ".inst 0x6e859580 // udot v0.4s, v12.16b, v5.16b\n"
+ "uzp1 v26.16b, v26.16b, v26.16b\n"
+ "uzp1 v6.16b, v6.16b, v6.16b\n"
+ "str s26, [x25, x27]\n"
+ "ldr q28, [%x[params], #0xe0]\n"
"uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9595d2 // udot v18.4s, v14.16b, v21.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "mov v22.16b, v0.16b\n .inst 0x6e899596 // udot v22.4s, v12.16b, v9.16b\n"
+ "uzp1 v6.16b, v6.16b, v6.16b\n"
"uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
- ".inst 0x6e8695d3 // udot v19.4s, v14.16b, v6.16b\n"
+ "str s6, [x24, x27]\n"
+ ".inst 0x6e889580 // udot v0.4s, v12.16b, v8.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "mov v30.16b, v5.16b\n"
- "str s20, [x21, x27]\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e869405 // udot v5.4s, v0.16b, v6.16b\n"
- ".inst 0x6e819419 // udot v25.4s, v0.16b, v1.16b\n"
- ".inst 0x6e819605 // udot v5.4s, v16.16b, v1.16b\n"
- "ext v6.16b, v6.16b, v6.16b, #0x1\n"
+ "str s25, [x23, x27]\n"
+ "mov v29.16b, v28.16b\n"
+ "str s20, [x22, x27]\n"
+ "mov v25.16b, v28.16b\n"
+ "mov v7.16b, v28.16b\n"
+ ".inst 0x6e88971c // udot v28.4s, v24.16b, v8.16b\n"
+ ".inst 0x6e8a9719 // udot v25.4s, v24.16b, v10.16b\n"
+ ".inst 0x6e8a97dc // udot v28.4s, v30.16b, v10.16b\n"
+ "ext v8.16b, v8.16b, v8.16b, #0x1\n"
"add x27, x27, #0x4\n"
- "ext v1.16b, v1.16b, v1.16b, #0x1\n"
+ "ext v10.16b, v10.16b, v10.16b, #0x1\n"
"movi v17.4s, #0x0\n"
- ".inst 0x6e86941e // udot v30.4s, v0.16b, v6.16b\n"
- ".inst 0x6e819414 // udot v20.4s, v0.16b, v1.16b\n"
- ".inst 0x6e8195d1 // udot v17.4s, v14.16b, v1.16b\n"
- ".inst 0x6e9a9619 // udot v25.4s, v16.16b, v26.16b\n"
- ".inst 0x6e9a97e5 // udot v5.4s, v31.16b, v26.16b\n"
- "ext v26.16b, v26.16b, v26.16b, #0x1\n"
- ".inst 0x6e81961e // udot v30.4s, v16.16b, v1.16b\n"
- "ldr q1, [x25, x28]\n"
- ".inst 0x6e9a9614 // udot v20.4s, v16.16b, v26.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9a95d1 // udot v17.4s, v14.16b, v26.16b\n"
- ".inst 0x6e9597f9 // udot v25.4s, v31.16b, v21.16b\n"
- "ext v21.16b, v21.16b, v21.16b, #0x1\n"
- ".inst 0x6e9a97fe // udot v30.4s, v31.16b, v26.16b\n"
- ".inst 0x6e9597f4 // udot v20.4s, v31.16b, v21.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9595d0 // udot v16.4s, v14.16b, v21.16b\n"
- ".inst 0x6e8695d1 // udot v17.4s, v14.16b, v6.16b\n"
- "ldr q6, [x12, x28]\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "ldr q9, [x15, x28]\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "ldp x15, x14, [%x[inptrs], #0x40]\n"
- "ldr q29, [x15, x28]\n"
- "ldr q28, [x14, x28]\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "ldp x13, x12, [%x[inptrs], #0x50]\n"
- "ldr q27, [x13, x28]\n"
- "ldr q26, [x12, x28]\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "ldr q16, [%x[params], #0x160]\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "ldr q31, [%x[params], #0x170]\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "ldr q0, [%x[params], #0x150]\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "ldr q4, [x10, x28]\n"
- "ldp x10, x9, [%x[inptrs], #0x60]\n"
- "ldr q24, [x10, x28]\n"
- "ldr q23, [x9, x28]\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "ldp x26, x25, [%x[inptrs], #0x70]\n"
- "ldr q22, [x26, x28]\n"
- "ldr q21, [x25, x28]\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smax v30.4s, v30.4s, v13.4s\n"
+ ".inst 0x6e88971d // udot v29.4s, v24.16b, v8.16b\n"
+ ".inst 0x6e8a9707 // udot v7.4s, v24.16b, v10.16b\n"
+ ".inst 0x6e8a9591 // udot v17.4s, v12.16b, v10.16b\n"
+ ".inst 0x6e8597d9 // udot v25.4s, v30.16b, v5.16b\n"
+ ".inst 0x6e85977c // udot v28.4s, v27.16b, v5.16b\n"
+ "ext v5.16b, v5.16b, v5.16b, #0x1\n"
+ ".inst 0x6e8a97dd // udot v29.4s, v30.16b, v10.16b\n"
+ "ldr q10, [x21, x28]\n"
+ ".inst 0x6e8597c7 // udot v7.4s, v30.16b, v5.16b\n"
+ "mls v28.4s, v0.4s, v16.4s\n"
+ ".inst 0x6e859591 // udot v17.4s, v12.16b, v5.16b\n"
+ ".inst 0x6e899779 // udot v25.4s, v27.16b, v9.16b\n"
+ "ext v9.16b, v9.16b, v9.16b, #0x1\n"
+ ".inst 0x6e85977d // udot v29.4s, v27.16b, v5.16b\n"
+ ".inst 0x6e899767 // udot v7.4s, v27.16b, v9.16b\n"
+ "sqrdmulh v28.4s, v28.4s, v15.4s\n"
+ "mov v18.16b, v17.16b\n .inst 0x6e899592 // udot v18.4s, v12.16b, v9.16b\n"
+ ".inst 0x6e889591 // udot v17.4s, v12.16b, v8.16b\n"
+ "ldr q8, [x12, x28]\n"
+ "mls v29.4s, v17.4s, v16.4s\n"
+ "mls v25.4s, v22.4s, v16.4s\n"
+ "mls v7.4s, v18.4s, v16.4s\n"
+ "and v17.16b, v28.16b, v23.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v29.4s, v29.4s, v15.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v15.4s\n"
+ "sqrdmulh v7.4s, v7.4s, v15.4s\n"
+ "ldr q15, [x15, x28]\n"
+ "sqadd v28.4s, v28.4s, v17.4s\n"
+ "ldp x21, x20, [%x[inptrs], #0x40]\n"
+ "ldr q22, [x21, x28]\n"
+ "ldr q3, [x20, x28]\n"
+ "and v24.16b, v29.16b, v23.16b\n"
+ "and v20.16b, v25.16b, v23.16b\n"
+ "and v17.16b, v7.16b, v23.16b\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "ldp x21, x20, [%x[inptrs], #0x50]\n"
+ "ldr q2, [x21, x28]\n"
+ "ldr q5, [x20, x28]\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "srshl v28.4s, v28.4s, v23.4s\n"
+ "sqadd v29.4s, v29.4s, v24.4s\n"
+ "ldr q6, [%x[params], #0x160]\n"
+ "sqadd v25.4s, v25.4s, v20.4s\n"
+ "ldr q20, [%x[params], #0x170]\n"
+ "sqadd v7.4s, v7.4s, v17.4s\n"
+ "ldr q1, [%x[params], #0x150]\n"
+ "add v28.4s, v28.4s, v14.4s\n"
+ "srshl v29.4s, v29.4s, v23.4s\n"
+ "srshl v25.4s, v25.4s, v23.4s\n"
+ "srshl v7.4s, v7.4s, v23.4s\n"
+ "ldr q26, [x10, x28]\n"
+ "ldp x21, x20, [%x[inptrs], #0x60]\n"
+ "ldr q27, [x21, x28]\n"
+ "ldr q30, [x20, x28]\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v25.4s, v25.4s, v14.4s\n"
+ "add v7.4s, v7.4s, v14.4s\n"
+ "ldp x21, x20, [%x[inptrs], #0x70]\n"
+ "ldr q23, [x21, x28]\n"
+ "ldr q9, [x20, x28]\n"
+ "smin v28.4s, v28.4s, v11.4s\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
"ldp x15, x14, [%x[inptrs], #0x0]\n"
"smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
+ "smax v7.4s, v7.4s, v13.4s\n"
"ldp x13, x12, [%x[inptrs], #0x10]\n"
"ldp x10, x9, [%x[inptrs], #0x20]\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "ldp x26, x25, [%x[inptrs], #0x30]\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s5, [x24, x27]\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v25.4s, v25.4s, v11.4s\n"
+ "ldp x26, x21, [%x[inptrs], #0x30]\n"
+ "smin v7.4s, v7.4s, v11.4s\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "str s28, [x25, x27]\n"
"uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "zip2 v5.16b, v9.16b, v7.16b\n"
- "zip1 v9.16b, v9.16b, v7.16b\n"
- "zip1 v7.16b, v8.16b, v6.16b\n"
- "zip2 v6.16b, v8.16b, v6.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v7.16b, v7.16b, v7.16b\n"
+ "zip2 v17.16b, v15.16b, v21.16b\n"
+ "zip1 v15.16b, v15.16b, v21.16b\n"
+ "zip1 v18.16b, v31.16b, v8.16b\n"
+ "zip2 v8.16b, v31.16b, v8.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
"uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "zip2 v8.16b, v9.16b, v7.16b\n"
- "str s20, [x21, x27]\n"
- "zip1 v9.16b, v9.16b, v7.16b\n"
- "zip1 v7.16b, v5.16b, v6.16b\n"
+ "str s29, [x24, x27]\n"
+ "uzp1 v7.16b, v7.16b, v7.16b\n"
+ "str s25, [x23, x27]\n"
+ "zip2 v25.16b, v15.16b, v18.16b\n"
+ "str s7, [x22, x27]\n"
+ "zip1 v15.16b, v15.16b, v18.16b\n"
+ "zip1 v7.16b, v17.16b, v8.16b\n"
"add x27, x27, #0x4\n"
- "zip2 v6.16b, v5.16b, v6.16b\n"
- "ldr q5, [%x[params], #0x140]\n"
- "zip2 v30.16b, v4.16b, v2.16b\n"
+ "zip2 v8.16b, v17.16b, v8.16b\n"
+ "ldr q31, [%x[params], #0x140]\n"
+ "zip2 v29.16b, v26.16b, v19.16b\n"
"add %x[params], %x[params], #0x180\n"
- "zip1 v4.16b, v4.16b, v2.16b\n"
- "zip1 v2.16b, v3.16b, v1.16b\n"
- "zip2 v1.16b, v3.16b, v1.16b\n"
- "zip2 v25.16b, v29.16b, v27.16b\n"
- "zip1 v29.16b, v29.16b, v27.16b\n"
- "zip1 v27.16b, v28.16b, v26.16b\n"
- "zip2 v26.16b, v28.16b, v26.16b\n"
- "zip2 v20.16b, v24.16b, v22.16b\n"
- "zip1 v24.16b, v24.16b, v22.16b\n"
- "zip1 v22.16b, v23.16b, v21.16b\n"
- "zip2 v21.16b, v23.16b, v21.16b\n"
- "zip2 v3.16b, v4.16b, v2.16b\n"
- "zip1 v4.16b, v4.16b, v2.16b\n"
- "zip1 v2.16b, v30.16b, v1.16b\n"
- "zip2 v1.16b, v30.16b, v1.16b\n"
- "zip2 v28.16b, v29.16b, v27.16b\n"
- "zip1 v29.16b, v29.16b, v27.16b\n"
- "zip1 v27.16b, v25.16b, v26.16b\n"
- "zip2 v26.16b, v25.16b, v26.16b\n"
- "zip2 v23.16b, v24.16b, v22.16b\n"
- "zip1 v24.16b, v24.16b, v22.16b\n"
- "zip1 v22.16b, v20.16b, v21.16b\n"
- "zip2 v21.16b, v20.16b, v21.16b\n"
- "mov v30.16b, v5.16b\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
+ "zip1 v26.16b, v26.16b, v19.16b\n"
+ "zip1 v28.16b, v4.16b, v10.16b\n"
+ "zip2 v10.16b, v4.16b, v10.16b\n"
+ "zip2 v24.16b, v22.16b, v2.16b\n"
+ "zip1 v22.16b, v22.16b, v2.16b\n"
+ "zip1 v21.16b, v3.16b, v5.16b\n"
+ "zip2 v5.16b, v3.16b, v5.16b\n"
+ "zip2 v18.16b, v27.16b, v23.16b\n"
+ "zip1 v27.16b, v27.16b, v23.16b\n"
+ "zip1 v17.16b, v30.16b, v9.16b\n"
+ "zip2 v9.16b, v30.16b, v9.16b\n"
+ "zip2 v23.16b, v26.16b, v28.16b\n"
+ "zip1 v26.16b, v26.16b, v28.16b\n"
+ "zip1 v3.16b, v29.16b, v10.16b\n"
+ "zip2 v10.16b, v29.16b, v10.16b\n"
+ "zip2 v19.16b, v22.16b, v21.16b\n"
+ "zip1 v22.16b, v22.16b, v21.16b\n"
+ "zip1 v0.16b, v24.16b, v5.16b\n"
+ "zip2 v5.16b, v24.16b, v5.16b\n"
+ "zip2 v24.16b, v27.16b, v17.16b\n"
+ "zip1 v27.16b, v27.16b, v17.16b\n"
+ "zip1 v2.16b, v18.16b, v9.16b\n"
+ "zip2 v9.16b, v18.16b, v9.16b\n"
+ "mov v30.16b, v31.16b\n"
+ "mov v29.16b, v31.16b\n"
+ "mov v28.16b, v31.16b\n"
"bgt 1b\n"
"2:" // Detached iteration
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8495d3 // udot v19.4s, v14.16b, v4.16b\n"
- ".inst 0x6e899405 // udot v5.4s, v0.16b, v9.16b\n"
+ "movi v21.4s, #0x0\n"
+ ".inst 0x6e9a9595 // udot v21.4s, v12.16b, v26.16b\n"
+ ".inst 0x6e8f943f // udot v31.4s, v1.16b, v15.16b\n"
"tst %x[n_channels], #0xf\n"
- ".inst 0x6e9d95d3 // udot v19.4s, v14.16b, v29.16b\n"
- ".inst 0x6e849419 // udot v25.4s, v0.16b, v4.16b\n"
+ ".inst 0x6e969595 // udot v21.4s, v12.16b, v22.16b\n"
+ ".inst 0x6e9a943d // udot v29.4s, v1.16b, v26.16b\n"
+ "movi v18.4s, #0x0\n"
"add x28, x28, #0x10\n"
- ".inst 0x6e849605 // udot v5.4s, v16.16b, v4.16b\n"
- "ext v4.16b, v4.16b, v4.16b, #0x1\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9895d2 // udot v18.4s, v14.16b, v24.16b\n"
- ".inst 0x6e8995d3 // udot v19.4s, v14.16b, v9.16b\n"
- "ext v9.16b, v9.16b, v9.16b, #0x1\n"
- ".inst 0x6e9d9619 // udot v25.4s, v16.16b, v29.16b\n"
- ".inst 0x6e9d97e5 // udot v5.4s, v31.16b, v29.16b\n"
- "ext v29.16b, v29.16b, v29.16b, #0x1\n"
- ".inst 0x6e89941e // udot v30.4s, v0.16b, v9.16b\n"
- ".inst 0x6e849414 // udot v20.4s, v0.16b, v4.16b\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x6e8495d1 // udot v17.4s, v14.16b, v4.16b\n"
- ".inst 0x6e9d95d1 // udot v17.4s, v14.16b, v29.16b\n"
- ".inst 0x6e9897f9 // udot v25.4s, v31.16b, v24.16b\n"
- "ext v24.16b, v24.16b, v24.16b, #0x1\n"
- ".inst 0x6e84961e // udot v30.4s, v16.16b, v4.16b\n"
+ ".inst 0x6e9a94df // udot v31.4s, v6.16b, v26.16b\n"
+ "ext v26.16b, v26.16b, v26.16b, #0x1\n"
+ "mov v17.16b, v21.16b\n .inst 0x6e9b9591 // udot v17.4s, v12.16b, v27.16b\n"
+ ".inst 0x6e8f9595 // udot v21.4s, v12.16b, v15.16b\n"
+ "ext v15.16b, v15.16b, v15.16b, #0x1\n"
+ ".inst 0x6e9a9592 // udot v18.4s, v12.16b, v26.16b\n"
+ ".inst 0x6e9694dd // udot v29.4s, v6.16b, v22.16b\n"
+ ".inst 0x6e96969f // udot v31.4s, v20.16b, v22.16b\n"
+ "ext v22.16b, v22.16b, v22.16b, #0x1\n"
+ ".inst 0x6e8f943e // udot v30.4s, v1.16b, v15.16b\n"
+ ".inst 0x6e9a943c // udot v28.4s, v1.16b, v26.16b\n"
+ "mls v31.4s, v21.4s, v16.4s\n"
+ ".inst 0x6e969592 // udot v18.4s, v12.16b, v22.16b\n"
+ ".inst 0x6e9b969d // udot v29.4s, v20.16b, v27.16b\n"
+ "ext v27.16b, v27.16b, v27.16b, #0x1\n"
+ ".inst 0x6e9a94de // udot v30.4s, v6.16b, v26.16b\n"
"ldr q4, [%x[params], #0x10]\n"
- ".inst 0x6e9d9614 // udot v20.4s, v16.16b, v29.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9895d0 // udot v16.4s, v14.16b, v24.16b\n"
- ".inst 0x6e8995d1 // udot v17.4s, v14.16b, v9.16b\n"
- "ldr q9, [%x[params], #0x0]\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- ".inst 0x6e9d97fe // udot v30.4s, v31.16b, v29.16b\n"
- ".inst 0x6e9897f4 // udot v20.4s, v31.16b, v24.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "ldr q9, [%x[params], #0x60]\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "ldr q16, [%x[params], #0x40]\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "ldr q31, [%x[params], #0x50]\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "ldr q0, [%x[params], #0x30]\n"
- "add v5.4s, v5.4s, v10.4s\n"
+ ".inst 0x6e9694dc // udot v28.4s, v6.16b, v22.16b\n"
+ "mls v29.4s, v17.4s, v16.4s\n"
+ "mov v21.16b, v18.16b\n .inst 0x6e9b9595 // udot v21.4s, v12.16b, v27.16b\n"
+ ".inst 0x6e8f9592 // udot v18.4s, v12.16b, v15.16b\n"
+ "ldr q17, [%x[params], #0x0]\n"
+ "sqrdmulh v31.4s, v31.4s, v17.4s\n"
+ ".inst 0x6e96969e // udot v30.4s, v20.16b, v22.16b\n"
+ ".inst 0x6e9b969c // udot v28.4s, v20.16b, v27.16b\n"
+ "mls v30.4s, v18.4s, v16.4s\n"
+ "mls v28.4s, v21.4s, v16.4s\n"
+ "and v27.16b, v31.16b, v4.16b\n"
+ "sshr v27.4s, v27.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v17.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v17.4s\n"
+ "sqrdmulh v28.4s, v28.4s, v17.4s\n"
+ "ldr q15, [%x[params], #0x60]\n"
+ "sqadd v31.4s, v31.4s, v27.4s\n"
+ "and v20.16b, v30.16b, v4.16b\n"
+ "and v18.16b, v29.16b, v4.16b\n"
+ "and v17.16b, v28.16b, v4.16b\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "srshl v31.4s, v31.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v20.4s\n"
+ "ldr q27, [%x[params], #0x40]\n"
+ "sqadd v29.4s, v29.4s, v18.4s\n"
+ "ldr q26, [%x[params], #0x50]\n"
+ "sqadd v28.4s, v28.4s, v17.4s\n"
+ "ldr q6, [%x[params], #0x30]\n"
+ "add v31.4s, v31.4s, v14.4s\n"
"srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
+ "srshl v29.4s, v29.4s, v4.4s\n"
+ "srshl v28.4s, v28.4s, v4.4s\n"
"ldr q4, [%x[params], #0x70]\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v14.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v28.4s, v28.4s, v14.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
"smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8395d3 // udot v19.4s, v14.16b, v3.16b\n"
- ".inst 0x6e9c95d3 // udot v19.4s, v14.16b, v28.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smin v30.4s, v30.4s, v11.4s\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v28.4s, v28.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
+ "movi v1.4s, #0x0\n"
+ ".inst 0x6e979581 // udot v1.4s, v12.16b, v23.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s5, [x24, x27]\n"
- "ldr q5, [%x[params], #0x20]\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9795d2 // udot v18.4s, v14.16b, v23.16b\n"
+ "str s31, [x25, x27]\n"
+ "ldr q31, [%x[params], #0x20]\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ ".inst 0x6e939581 // udot v1.4s, v12.16b, v19.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
- ".inst 0x6e8895d3 // udot v19.4s, v14.16b, v8.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "mov v30.16b, v5.16b\n"
- "str s20, [x21, x27]\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e889405 // udot v5.4s, v0.16b, v8.16b\n"
- ".inst 0x6e839419 // udot v25.4s, v0.16b, v3.16b\n"
- ".inst 0x6e839605 // udot v5.4s, v16.16b, v3.16b\n"
- "ext v8.16b, v8.16b, v8.16b, #0x1\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "str s30, [x24, x27]\n"
+ "mov v22.16b, v1.16b\n .inst 0x6e989596 // udot v22.4s, v12.16b, v24.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "str s29, [x23, x27]\n"
+ "mov v29.16b, v31.16b\n"
+ ".inst 0x6e999581 // udot v1.4s, v12.16b, v25.16b\n"
+ "str s28, [x22, x27]\n"
+ "mov v21.16b, v31.16b\n"
+ "mov v20.16b, v31.16b\n"
+ ".inst 0x6e9994df // udot v31.4s, v6.16b, v25.16b\n"
+ ".inst 0x6e9794d5 // udot v21.4s, v6.16b, v23.16b\n"
+ ".inst 0x6e97977f // udot v31.4s, v27.16b, v23.16b\n"
+ "ext v25.16b, v25.16b, v25.16b, #0x1\n"
"add x27, x27, #0x4\n"
- "ext v3.16b, v3.16b, v3.16b, #0x1\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x6e88941e // udot v30.4s, v0.16b, v8.16b\n"
- ".inst 0x6e839414 // udot v20.4s, v0.16b, v3.16b\n"
- ".inst 0x6e8395d1 // udot v17.4s, v14.16b, v3.16b\n"
- ".inst 0x6e9c9619 // udot v25.4s, v16.16b, v28.16b\n"
- ".inst 0x6e9c97e5 // udot v5.4s, v31.16b, v28.16b\n"
- "ext v28.16b, v28.16b, v28.16b, #0x1\n"
- ".inst 0x6e83961e // udot v30.4s, v16.16b, v3.16b\n"
- ".inst 0x6e9c9614 // udot v20.4s, v16.16b, v28.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9c95d1 // udot v17.4s, v14.16b, v28.16b\n"
- ".inst 0x6e9797f9 // udot v25.4s, v31.16b, v23.16b\n"
"ext v23.16b, v23.16b, v23.16b, #0x1\n"
- ".inst 0x6e9c97fe // udot v30.4s, v31.16b, v28.16b\n"
- ".inst 0x6e9797f4 // udot v20.4s, v31.16b, v23.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9795d0 // udot v16.4s, v14.16b, v23.16b\n"
- ".inst 0x6e8895d1 // udot v17.4s, v14.16b, v8.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "ldr q9, [%x[params], #0xc0]\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "ldr q16, [%x[params], #0xa0]\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "ldr q31, [%x[params], #0xb0]\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "ldr q0, [%x[params], #0x90]\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
+ "movi v18.4s, #0x0\n"
+ ".inst 0x6e9994dd // udot v29.4s, v6.16b, v25.16b\n"
+ ".inst 0x6e9794d4 // udot v20.4s, v6.16b, v23.16b\n"
+ ".inst 0x6e979592 // udot v18.4s, v12.16b, v23.16b\n"
+ ".inst 0x6e939775 // udot v21.4s, v27.16b, v19.16b\n"
+ ".inst 0x6e93975f // udot v31.4s, v26.16b, v19.16b\n"
+ "ext v19.16b, v19.16b, v19.16b, #0x1\n"
+ ".inst 0x6e97977d // udot v29.4s, v27.16b, v23.16b\n"
+ ".inst 0x6e939774 // udot v20.4s, v27.16b, v19.16b\n"
+ "mls v31.4s, v1.4s, v16.4s\n"
+ ".inst 0x6e939592 // udot v18.4s, v12.16b, v19.16b\n"
+ ".inst 0x6e989755 // udot v21.4s, v26.16b, v24.16b\n"
+ "ext v24.16b, v24.16b, v24.16b, #0x1\n"
+ ".inst 0x6e93975d // udot v29.4s, v26.16b, v19.16b\n"
+ ".inst 0x6e989754 // udot v20.4s, v26.16b, v24.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v15.4s\n"
+ "mov v17.16b, v18.16b\n .inst 0x6e989591 // udot v17.4s, v12.16b, v24.16b\n"
+ ".inst 0x6e999592 // udot v18.4s, v12.16b, v25.16b\n"
+ "mls v29.4s, v18.4s, v16.4s\n"
+ "mls v21.4s, v22.4s, v16.4s\n"
+ "mls v20.4s, v17.4s, v16.4s\n"
+ "and v17.16b, v31.16b, v4.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v29.4s, v29.4s, v15.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v15.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v15.4s\n"
+ "ldr q27, [%x[params], #0xc0]\n"
+ "sqadd v31.4s, v31.4s, v17.4s\n"
+ "and v19.16b, v29.16b, v4.16b\n"
+ "and v18.16b, v21.16b, v4.16b\n"
+ "and v17.16b, v20.16b, v4.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "srshl v31.4s, v31.4s, v4.4s\n"
+ "sqadd v29.4s, v29.4s, v19.4s\n"
+ "ldr q26, [%x[params], #0xa0]\n"
+ "sqadd v21.4s, v21.4s, v18.4s\n"
+ "ldr q25, [%x[params], #0xb0]\n"
+ "sqadd v20.4s, v20.4s, v17.4s\n"
+ "ldr q24, [%x[params], #0x90]\n"
+ "add v31.4s, v31.4s, v14.4s\n"
+ "srshl v29.4s, v29.4s, v4.4s\n"
+ "srshl v21.4s, v21.4s, v4.4s\n"
"srshl v20.4s, v20.4s, v4.4s\n"
- "ldr q4, [%x[params], #0xd0]\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
+ "ldr q1, [%x[params], #0xd0]\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v21.4s, v21.4s, v14.4s\n"
+ "add v20.4s, v20.4s, v14.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v21.4s, v21.4s, v13.4s\n"
"smax v20.4s, v20.4s, v13.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8295d3 // udot v19.4s, v14.16b, v2.16b\n"
- ".inst 0x6e9b95d3 // udot v19.4s, v14.16b, v27.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s5, [x24, x27]\n"
- "ldr q5, [%x[params], #0x80]\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v21.4s, v21.4s, v11.4s\n"
+ "smin v20.4s, v20.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
+ "movi v23.4s, #0x0\n"
+ ".inst 0x6e839597 // udot v23.4s, v12.16b, v3.16b\n"
+ ".inst 0x6e809597 // udot v23.4s, v12.16b, v0.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "str s31, [x25, x27]\n"
+ "ldr q31, [%x[params], #0x80]\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9695d2 // udot v18.4s, v14.16b, v22.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
- ".inst 0x6e8795d3 // udot v19.4s, v14.16b, v7.16b\n"
+ "mov v22.16b, v23.16b\n .inst 0x6e829596 // udot v22.4s, v12.16b, v2.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
+ "str s29, [x24, x27]\n"
+ ".inst 0x6e879597 // udot v23.4s, v12.16b, v7.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "mov v30.16b, v5.16b\n"
- "str s20, [x21, x27]\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e879405 // udot v5.4s, v0.16b, v7.16b\n"
- ".inst 0x6e829419 // udot v25.4s, v0.16b, v2.16b\n"
- ".inst 0x6e829605 // udot v5.4s, v16.16b, v2.16b\n"
+ "str s21, [x23, x27]\n"
+ "mov v21.16b, v31.16b\n"
+ "str s20, [x22, x27]\n"
+ "mov v4.16b, v31.16b\n"
+ "mov v20.16b, v31.16b\n"
+ ".inst 0x6e87971f // udot v31.4s, v24.16b, v7.16b\n"
+ ".inst 0x6e839704 // udot v4.4s, v24.16b, v3.16b\n"
+ ".inst 0x6e83975f // udot v31.4s, v26.16b, v3.16b\n"
"ext v7.16b, v7.16b, v7.16b, #0x1\n"
"add x27, x27, #0x4\n"
+ "ext v3.16b, v3.16b, v3.16b, #0x1\n"
+ "movi v18.4s, #0x0\n"
+ ".inst 0x6e879715 // udot v21.4s, v24.16b, v7.16b\n"
+ ".inst 0x6e839714 // udot v20.4s, v24.16b, v3.16b\n"
+ ".inst 0x6e839592 // udot v18.4s, v12.16b, v3.16b\n"
+ ".inst 0x6e809744 // udot v4.4s, v26.16b, v0.16b\n"
+ ".inst 0x6e80973f // udot v31.4s, v25.16b, v0.16b\n"
+ "ext v0.16b, v0.16b, v0.16b, #0x1\n"
+ ".inst 0x6e839755 // udot v21.4s, v26.16b, v3.16b\n"
+ ".inst 0x6e809754 // udot v20.4s, v26.16b, v0.16b\n"
+ "mls v31.4s, v23.4s, v16.4s\n"
+ ".inst 0x6e809592 // udot v18.4s, v12.16b, v0.16b\n"
+ ".inst 0x6e829724 // udot v4.4s, v25.16b, v2.16b\n"
"ext v2.16b, v2.16b, v2.16b, #0x1\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x6e87941e // udot v30.4s, v0.16b, v7.16b\n"
- ".inst 0x6e829414 // udot v20.4s, v0.16b, v2.16b\n"
- ".inst 0x6e8295d1 // udot v17.4s, v14.16b, v2.16b\n"
- ".inst 0x6e9b9619 // udot v25.4s, v16.16b, v27.16b\n"
- ".inst 0x6e9b97e5 // udot v5.4s, v31.16b, v27.16b\n"
- "ext v27.16b, v27.16b, v27.16b, #0x1\n"
- ".inst 0x6e82961e // udot v30.4s, v16.16b, v2.16b\n"
- ".inst 0x6e9b9614 // udot v20.4s, v16.16b, v27.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9b95d1 // udot v17.4s, v14.16b, v27.16b\n"
- ".inst 0x6e9697f9 // udot v25.4s, v31.16b, v22.16b\n"
- "ext v22.16b, v22.16b, v22.16b, #0x1\n"
- ".inst 0x6e9b97fe // udot v30.4s, v31.16b, v27.16b\n"
- ".inst 0x6e9697f4 // udot v20.4s, v31.16b, v22.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9695d0 // udot v16.4s, v14.16b, v22.16b\n"
- ".inst 0x6e8795d1 // udot v17.4s, v14.16b, v7.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "ldr q9, [%x[params], #0x120]\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "ldr q16, [%x[params], #0x100]\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "ldr q31, [%x[params], #0x110]\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "ldr q0, [%x[params], #0xf0]\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "ldr q4, [%x[params], #0x130]\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
+ ".inst 0x6e809735 // udot v21.4s, v25.16b, v0.16b\n"
+ ".inst 0x6e829734 // udot v20.4s, v25.16b, v2.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v27.4s\n"
+ "mov v17.16b, v18.16b\n .inst 0x6e829591 // udot v17.4s, v12.16b, v2.16b\n"
+ ".inst 0x6e879592 // udot v18.4s, v12.16b, v7.16b\n"
+ "mls v21.4s, v18.4s, v16.4s\n"
+ "mls v4.4s, v22.4s, v16.4s\n"
+ "mls v20.4s, v17.4s, v16.4s\n"
+ "and v17.16b, v31.16b, v1.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v21.4s, v21.4s, v27.4s\n"
+ "sqrdmulh v4.4s, v4.4s, v27.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v27.4s\n"
+ "ldr q30, [%x[params], #0x120]\n"
+ "sqadd v31.4s, v31.4s, v17.4s\n"
+ "and v19.16b, v21.16b, v1.16b\n"
+ "and v18.16b, v4.16b, v1.16b\n"
+ "and v17.16b, v20.16b, v1.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "srshl v31.4s, v31.4s, v1.4s\n"
+ "sqadd v21.4s, v21.4s, v19.4s\n"
+ "ldr q29, [%x[params], #0x100]\n"
+ "sqadd v4.4s, v4.4s, v18.4s\n"
+ "ldr q28, [%x[params], #0x110]\n"
+ "sqadd v20.4s, v20.4s, v17.4s\n"
+ "ldr q27, [%x[params], #0xf0]\n"
+ "add v31.4s, v31.4s, v14.4s\n"
+ "srshl v21.4s, v21.4s, v1.4s\n"
+ "srshl v4.4s, v4.4s, v1.4s\n"
+ "srshl v20.4s, v20.4s, v1.4s\n"
+ "ldr q26, [%x[params], #0x130]\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
+ "add v21.4s, v21.4s, v14.4s\n"
+ "add v4.4s, v4.4s, v14.4s\n"
+ "add v20.4s, v20.4s, v14.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
+ "smax v21.4s, v21.4s, v13.4s\n"
+ "smax v4.4s, v4.4s, v13.4s\n"
"smax v20.4s, v20.4s, v13.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8195d3 // udot v19.4s, v14.16b, v1.16b\n"
- ".inst 0x6e9a95d3 // udot v19.4s, v14.16b, v26.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s5, [x24, x27]\n"
- "ldr q5, [%x[params], #0xe0]\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "smin v21.4s, v21.4s, v11.4s\n"
+ "smin v4.4s, v4.4s, v11.4s\n"
+ "smin v20.4s, v20.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
+ "movi v25.4s, #0x0\n"
+ ".inst 0x6e8a9599 // udot v25.4s, v12.16b, v10.16b\n"
+ ".inst 0x6e859599 // udot v25.4s, v12.16b, v5.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
+ "str s31, [x25, x27]\n"
+ "ldr q24, [%x[params], #0xe0]\n"
+ "uzp1 v4.16b, v4.16b, v4.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9595d2 // udot v18.4s, v14.16b, v21.16b\n"
+ "mov v23.16b, v25.16b\n .inst 0x6e899597 // udot v23.4s, v12.16b, v9.16b\n"
"add %x[params], %x[params], #0x140\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
- ".inst 0x6e8695d3 // udot v19.4s, v14.16b, v6.16b\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
+ "uzp1 v4.16b, v4.16b, v4.16b\n"
+ "str s21, [x24, x27]\n"
+ ".inst 0x6e889599 // udot v25.4s, v12.16b, v8.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "mov v30.16b, v5.16b\n"
- "str s20, [x21, x27]\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e869405 // udot v5.4s, v0.16b, v6.16b\n"
- ".inst 0x6e819419 // udot v25.4s, v0.16b, v1.16b\n"
- ".inst 0x6e819605 // udot v5.4s, v16.16b, v1.16b\n"
- "ext v6.16b, v6.16b, v6.16b, #0x1\n"
+ "str s4, [x23, x27]\n"
+ "mov v22.16b, v24.16b\n"
+ "str s20, [x22, x27]\n"
+ "mov v21.16b, v24.16b\n"
+ "mov v20.16b, v24.16b\n"
+ ".inst 0x6e889778 // udot v24.4s, v27.16b, v8.16b\n"
+ ".inst 0x6e8a9775 // udot v21.4s, v27.16b, v10.16b\n"
+ ".inst 0x6e8a97b8 // udot v24.4s, v29.16b, v10.16b\n"
+ "ext v8.16b, v8.16b, v8.16b, #0x1\n"
"add x27, x27, #0x4\n"
- "ext v1.16b, v1.16b, v1.16b, #0x1\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x6e86941e // udot v30.4s, v0.16b, v6.16b\n"
- ".inst 0x6e819414 // udot v20.4s, v0.16b, v1.16b\n"
- ".inst 0x6e8195d1 // udot v17.4s, v14.16b, v1.16b\n"
- ".inst 0x6e9a9619 // udot v25.4s, v16.16b, v26.16b\n"
- ".inst 0x6e9a97e5 // udot v5.4s, v31.16b, v26.16b\n"
- "ext v26.16b, v26.16b, v26.16b, #0x1\n"
- ".inst 0x6e81961e // udot v30.4s, v16.16b, v1.16b\n"
- ".inst 0x6e9a9614 // udot v20.4s, v16.16b, v26.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9a95d1 // udot v17.4s, v14.16b, v26.16b\n"
- ".inst 0x6e9597f9 // udot v25.4s, v31.16b, v21.16b\n"
- "ext v21.16b, v21.16b, v21.16b, #0x1\n"
- ".inst 0x6e9a97fe // udot v30.4s, v31.16b, v26.16b\n"
- ".inst 0x6e9597f4 // udot v20.4s, v31.16b, v21.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9595d0 // udot v16.4s, v14.16b, v21.16b\n"
- ".inst 0x6e8695d1 // udot v17.4s, v14.16b, v6.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smax v5.4s, v5.4s, v13.4s\n"
- "smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
+ "ext v10.16b, v10.16b, v10.16b, #0x1\n"
+ "movi v18.4s, #0x0\n"
+ ".inst 0x6e889776 // udot v22.4s, v27.16b, v8.16b\n"
+ ".inst 0x6e8a9774 // udot v20.4s, v27.16b, v10.16b\n"
+ ".inst 0x6e8a9592 // udot v18.4s, v12.16b, v10.16b\n"
+ ".inst 0x6e8597b5 // udot v21.4s, v29.16b, v5.16b\n"
+ ".inst 0x6e859798 // udot v24.4s, v28.16b, v5.16b\n"
+ "ext v5.16b, v5.16b, v5.16b, #0x1\n"
+ ".inst 0x6e8a97b6 // udot v22.4s, v29.16b, v10.16b\n"
+ ".inst 0x6e8597b4 // udot v20.4s, v29.16b, v5.16b\n"
+ "mls v24.4s, v25.4s, v16.4s\n"
+ ".inst 0x6e859592 // udot v18.4s, v12.16b, v5.16b\n"
+ ".inst 0x6e899795 // udot v21.4s, v28.16b, v9.16b\n"
+ "ext v9.16b, v9.16b, v9.16b, #0x1\n"
+ ".inst 0x6e859796 // udot v22.4s, v28.16b, v5.16b\n"
+ ".inst 0x6e899794 // udot v20.4s, v28.16b, v9.16b\n"
+ "sqrdmulh v24.4s, v24.4s, v30.4s\n"
+ "mov v17.16b, v18.16b\n .inst 0x6e899591 // udot v17.4s, v12.16b, v9.16b\n"
+ ".inst 0x6e889592 // udot v18.4s, v12.16b, v8.16b\n"
+ "mls v22.4s, v18.4s, v16.4s\n"
+ "mls v21.4s, v23.4s, v16.4s\n"
+ "mls v20.4s, v17.4s, v16.4s\n"
+ "and v17.16b, v24.16b, v26.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v22.4s, v22.4s, v30.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v30.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v30.4s\n"
+ "sqadd v24.4s, v24.4s, v17.4s\n"
+ "and v19.16b, v22.16b, v26.16b\n"
+ "and v18.16b, v21.16b, v26.16b\n"
+ "and v17.16b, v20.16b, v26.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqadd v22.4s, v22.4s, v19.4s\n"
+ "sqadd v21.4s, v21.4s, v18.4s\n"
+ "sqadd v20.4s, v20.4s, v17.4s\n"
+ "srshl v24.4s, v24.4s, v26.4s\n"
+ "srshl v22.4s, v22.4s, v26.4s\n"
+ "srshl v21.4s, v21.4s, v26.4s\n"
+ "srshl v20.4s, v20.4s, v26.4s\n"
+ "add v24.4s, v24.4s, v14.4s\n"
+ "add v22.4s, v22.4s, v14.4s\n"
+ "add v21.4s, v21.4s, v14.4s\n"
+ "add v20.4s, v20.4s, v14.4s\n"
+ "smax v24.4s, v24.4s, v13.4s\n"
+ "smax v22.4s, v22.4s, v13.4s\n"
+ "smax v21.4s, v21.4s, v13.4s\n"
"smax v20.4s, v20.4s, v13.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "smin v24.4s, v24.4s, v11.4s\n"
+ "smin v22.4s, v22.4s, v11.4s\n"
+ "smin v21.4s, v21.4s, v11.4s\n"
+ "smin v20.4s, v20.4s, v11.4s\n"
+ "uzp1 v24.16b, v24.16b, v24.16b\n"
+ "uzp1 v22.16b, v22.16b, v22.16b\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
- "str s5, [x24, x27]\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s30, [x23, x27]\n"
+ "uzp1 v24.16b, v24.16b, v24.16b\n"
+ "str s24, [x25, x27]\n"
+ "uzp1 v22.16b, v22.16b, v22.16b\n"
+ "uzp1 v21.16b, v21.16b, v21.16b\n"
+ "str s22, [x24, x27]\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s25, [x22, x27]\n"
- "str s20, [x21, x27]\n"
+ "str s21, [x23, x27]\n"
+ "str s20, [x22, x27]\n"
"add x27, x27, #0x4\n"
"beq 35f\n"
"3:" // Oddments
@@ -869,794 +861,794 @@ void a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst_impl(
"add x10, x10, x28\n"
"add x9, x9, x28\n"
"add x26, x26, x28\n"
- "add x25, x25, x28\n"
+ "add x21, x21, x28\n"
"tbz %x[n_channels], #3, 7f\n"
- "ldr d9, [x15], #0x8\n"
- "ldr d8, [x14], #0x8\n"
+ "ldr d15, [x15], #0x8\n"
+ "ldr d25, [x14], #0x8\n"
"ldr d7, [x13], #0x8\n"
- "ldr d6, [x12], #0x8\n"
- "ldr d4, [x10], #0x8\n"
- "ldr d3, [x9], #0x8\n"
- "ldr d2, [x26], #0x8\n"
- "ldr d1, [x25], #0x8\n"
+ "ldr d8, [x12], #0x8\n"
+ "ldr d26, [x10], #0x8\n"
+ "ldr d23, [x9], #0x8\n"
+ "ldr d3, [x26], #0x8\n"
+ "ldr d10, [x21], #0x8\n"
"tbz %x[n_channels], #2, 5f\n"
- "ld1 { v9.s }[2], [x15], #0x4\n"
- "ld1 { v8.s }[2], [x14], #0x4\n"
+ "ld1 { v15.s }[2], [x15], #0x4\n"
+ "ld1 { v25.s }[2], [x14], #0x4\n"
"ld1 { v7.s }[2], [x13], #0x4\n"
- "ld1 { v6.s }[2], [x12], #0x4\n"
- "ld1 { v4.s }[2], [x10], #0x4\n"
- "ld1 { v3.s }[2], [x9], #0x4\n"
- "ld1 { v2.s }[2], [x26], #0x4\n"
- "ld1 { v1.s }[2], [x25], #0x4\n"
+ "ld1 { v8.s }[2], [x12], #0x4\n"
+ "ld1 { v26.s }[2], [x10], #0x4\n"
+ "ld1 { v23.s }[2], [x9], #0x4\n"
+ "ld1 { v3.s }[2], [x26], #0x4\n"
+ "ld1 { v10.s }[2], [x21], #0x4\n"
"tbz %x[n_channels], #1, 4f\n"
- "ld1 { v9.h }[6], [x15], #0x2\n"
- "ld1 { v8.h }[6], [x14], #0x2\n"
+ "ld1 { v15.h }[6], [x15], #0x2\n"
+ "ld1 { v25.h }[6], [x14], #0x2\n"
"ld1 { v7.h }[6], [x13], #0x2\n"
- "ld1 { v6.h }[6], [x12], #0x2\n"
- "ld1 { v4.h }[6], [x10], #0x2\n"
- "ld1 { v3.h }[6], [x9], #0x2\n"
- "ld1 { v2.h }[6], [x26], #0x2\n"
- "ld1 { v1.h }[6], [x25], #0x2\n"
+ "ld1 { v8.h }[6], [x12], #0x2\n"
+ "ld1 { v26.h }[6], [x10], #0x2\n"
+ "ld1 { v23.h }[6], [x9], #0x2\n"
+ "ld1 { v3.h }[6], [x26], #0x2\n"
+ "ld1 { v10.h }[6], [x21], #0x2\n"
"tbz %x[n_channels], #0, 11f\n"
- "ld1 { v9.b }[14], [x15], #0x1\n"
- "ld1 { v8.b }[14], [x14], #0x1\n"
+ "ld1 { v15.b }[14], [x15], #0x1\n"
+ "ld1 { v25.b }[14], [x14], #0x1\n"
"ld1 { v7.b }[14], [x13], #0x1\n"
- "ld1 { v6.b }[14], [x12], #0x1\n"
- "ld1 { v4.b }[14], [x10], #0x1\n"
- "ld1 { v3.b }[14], [x9], #0x1\n"
- "ld1 { v2.b }[14], [x26], #0x1\n"
- "ld1 { v1.b }[14], [x25], #0x1\n"
+ "ld1 { v8.b }[14], [x12], #0x1\n"
+ "ld1 { v26.b }[14], [x10], #0x1\n"
+ "ld1 { v23.b }[14], [x9], #0x1\n"
+ "ld1 { v3.b }[14], [x26], #0x1\n"
+ "ld1 { v10.b }[14], [x21], #0x1\n"
"b 11f\n"
"4:" // Oddments: Load (A): Bit 3: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 11f\n"
- "ld1 { v9.b }[12], [x15], #0x1\n"
- "ld1 { v8.b }[12], [x14], #0x1\n"
+ "ld1 { v15.b }[12], [x15], #0x1\n"
+ "ld1 { v25.b }[12], [x14], #0x1\n"
"ld1 { v7.b }[12], [x13], #0x1\n"
- "ld1 { v6.b }[12], [x12], #0x1\n"
- "ld1 { v4.b }[12], [x10], #0x1\n"
- "ld1 { v3.b }[12], [x9], #0x1\n"
- "ld1 { v2.b }[12], [x26], #0x1\n"
- "ld1 { v1.b }[12], [x25], #0x1\n"
+ "ld1 { v8.b }[12], [x12], #0x1\n"
+ "ld1 { v26.b }[12], [x10], #0x1\n"
+ "ld1 { v23.b }[12], [x9], #0x1\n"
+ "ld1 { v3.b }[12], [x26], #0x1\n"
+ "ld1 { v10.b }[12], [x21], #0x1\n"
"b 11f\n"
"5:" // Oddments: Load (A): Bit 3: Bit 2: Unset
"tbz %x[n_channels], #1, 6f\n"
- "ld1 { v9.h }[4], [x15], #0x2\n"
- "ld1 { v8.h }[4], [x14], #0x2\n"
+ "ld1 { v15.h }[4], [x15], #0x2\n"
+ "ld1 { v25.h }[4], [x14], #0x2\n"
"ld1 { v7.h }[4], [x13], #0x2\n"
- "ld1 { v6.h }[4], [x12], #0x2\n"
- "ld1 { v4.h }[4], [x10], #0x2\n"
- "ld1 { v3.h }[4], [x9], #0x2\n"
- "ld1 { v2.h }[4], [x26], #0x2\n"
- "ld1 { v1.h }[4], [x25], #0x2\n"
+ "ld1 { v8.h }[4], [x12], #0x2\n"
+ "ld1 { v26.h }[4], [x10], #0x2\n"
+ "ld1 { v23.h }[4], [x9], #0x2\n"
+ "ld1 { v3.h }[4], [x26], #0x2\n"
+ "ld1 { v10.h }[4], [x21], #0x2\n"
"tbz %x[n_channels], #0, 11f\n"
- "ld1 { v9.b }[10], [x15], #0x1\n"
- "ld1 { v8.b }[10], [x14], #0x1\n"
+ "ld1 { v15.b }[10], [x15], #0x1\n"
+ "ld1 { v25.b }[10], [x14], #0x1\n"
"ld1 { v7.b }[10], [x13], #0x1\n"
- "ld1 { v6.b }[10], [x12], #0x1\n"
- "ld1 { v4.b }[10], [x10], #0x1\n"
- "ld1 { v3.b }[10], [x9], #0x1\n"
- "ld1 { v2.b }[10], [x26], #0x1\n"
- "ld1 { v1.b }[10], [x25], #0x1\n"
+ "ld1 { v8.b }[10], [x12], #0x1\n"
+ "ld1 { v26.b }[10], [x10], #0x1\n"
+ "ld1 { v23.b }[10], [x9], #0x1\n"
+ "ld1 { v3.b }[10], [x26], #0x1\n"
+ "ld1 { v10.b }[10], [x21], #0x1\n"
"b 11f\n"
"6:" // Oddments: Load (A): Bit 3: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 11f\n"
- "ld1 { v9.b }[8], [x15], #0x1\n"
- "ld1 { v8.b }[8], [x14], #0x1\n"
+ "ld1 { v15.b }[8], [x15], #0x1\n"
+ "ld1 { v25.b }[8], [x14], #0x1\n"
"ld1 { v7.b }[8], [x13], #0x1\n"
- "ld1 { v6.b }[8], [x12], #0x1\n"
- "ld1 { v4.b }[8], [x10], #0x1\n"
- "ld1 { v3.b }[8], [x9], #0x1\n"
- "ld1 { v2.b }[8], [x26], #0x1\n"
- "ld1 { v1.b }[8], [x25], #0x1\n"
+ "ld1 { v8.b }[8], [x12], #0x1\n"
+ "ld1 { v26.b }[8], [x10], #0x1\n"
+ "ld1 { v23.b }[8], [x9], #0x1\n"
+ "ld1 { v3.b }[8], [x26], #0x1\n"
+ "ld1 { v10.b }[8], [x21], #0x1\n"
"b 11f\n"
"7:" // Oddments: Load (A): Bit 3: Unset
"tbz %x[n_channels], #2, 9f\n"
- "ldr s9, [x15], #0x4\n"
- "ldr s8, [x14], #0x4\n"
+ "ldr s15, [x15], #0x4\n"
+ "ldr s25, [x14], #0x4\n"
"ldr s7, [x13], #0x4\n"
- "ldr s6, [x12], #0x4\n"
- "ldr s4, [x10], #0x4\n"
- "ldr s3, [x9], #0x4\n"
- "ldr s2, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s8, [x12], #0x4\n"
+ "ldr s26, [x10], #0x4\n"
+ "ldr s23, [x9], #0x4\n"
+ "ldr s3, [x26], #0x4\n"
+ "ldr s10, [x21], #0x4\n"
"tbz %x[n_channels], #1, 8f\n"
- "ld1 { v9.h }[2], [x15], #0x2\n"
- "ld1 { v8.h }[2], [x14], #0x2\n"
+ "ld1 { v15.h }[2], [x15], #0x2\n"
+ "ld1 { v25.h }[2], [x14], #0x2\n"
"ld1 { v7.h }[2], [x13], #0x2\n"
- "ld1 { v6.h }[2], [x12], #0x2\n"
- "ld1 { v4.h }[2], [x10], #0x2\n"
- "ld1 { v3.h }[2], [x9], #0x2\n"
- "ld1 { v2.h }[2], [x26], #0x2\n"
- "ld1 { v1.h }[2], [x25], #0x2\n"
+ "ld1 { v8.h }[2], [x12], #0x2\n"
+ "ld1 { v26.h }[2], [x10], #0x2\n"
+ "ld1 { v23.h }[2], [x9], #0x2\n"
+ "ld1 { v3.h }[2], [x26], #0x2\n"
+ "ld1 { v10.h }[2], [x21], #0x2\n"
"tbz %x[n_channels], #0, 11f\n"
- "ld1 { v9.b }[6], [x15], #0x1\n"
- "ld1 { v8.b }[6], [x14], #0x1\n"
+ "ld1 { v15.b }[6], [x15], #0x1\n"
+ "ld1 { v25.b }[6], [x14], #0x1\n"
"ld1 { v7.b }[6], [x13], #0x1\n"
- "ld1 { v6.b }[6], [x12], #0x1\n"
- "ld1 { v4.b }[6], [x10], #0x1\n"
- "ld1 { v3.b }[6], [x9], #0x1\n"
- "ld1 { v2.b }[6], [x26], #0x1\n"
- "ld1 { v1.b }[6], [x25], #0x1\n"
+ "ld1 { v8.b }[6], [x12], #0x1\n"
+ "ld1 { v26.b }[6], [x10], #0x1\n"
+ "ld1 { v23.b }[6], [x9], #0x1\n"
+ "ld1 { v3.b }[6], [x26], #0x1\n"
+ "ld1 { v10.b }[6], [x21], #0x1\n"
"b 11f\n"
"8:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 11f\n"
- "ld1 { v9.b }[4], [x15], #0x1\n"
- "ld1 { v8.b }[4], [x14], #0x1\n"
+ "ld1 { v15.b }[4], [x15], #0x1\n"
+ "ld1 { v25.b }[4], [x14], #0x1\n"
"ld1 { v7.b }[4], [x13], #0x1\n"
- "ld1 { v6.b }[4], [x12], #0x1\n"
- "ld1 { v4.b }[4], [x10], #0x1\n"
- "ld1 { v3.b }[4], [x9], #0x1\n"
- "ld1 { v2.b }[4], [x26], #0x1\n"
- "ld1 { v1.b }[4], [x25], #0x1\n"
+ "ld1 { v8.b }[4], [x12], #0x1\n"
+ "ld1 { v26.b }[4], [x10], #0x1\n"
+ "ld1 { v23.b }[4], [x9], #0x1\n"
+ "ld1 { v3.b }[4], [x26], #0x1\n"
+ "ld1 { v10.b }[4], [x21], #0x1\n"
"b 11f\n"
"9:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Unset
"tbz %x[n_channels], #1, 10f\n"
- "ldr h9, [x15], #0x2\n"
- "ldr h8, [x14], #0x2\n"
+ "ldr h15, [x15], #0x2\n"
+ "ldr h25, [x14], #0x2\n"
"ldr h7, [x13], #0x2\n"
- "ldr h6, [x12], #0x2\n"
- "ldr h4, [x10], #0x2\n"
- "ldr h3, [x9], #0x2\n"
- "ldr h2, [x26], #0x2\n"
- "ldr h1, [x25], #0x2\n"
+ "ldr h8, [x12], #0x2\n"
+ "ldr h26, [x10], #0x2\n"
+ "ldr h23, [x9], #0x2\n"
+ "ldr h3, [x26], #0x2\n"
+ "ldr h10, [x21], #0x2\n"
"tbz %x[n_channels], #0, 11f\n"
- "ld1 { v9.b }[2], [x15], #0x1\n"
- "ld1 { v8.b }[2], [x14], #0x1\n"
+ "ld1 { v15.b }[2], [x15], #0x1\n"
+ "ld1 { v25.b }[2], [x14], #0x1\n"
"ld1 { v7.b }[2], [x13], #0x1\n"
- "ld1 { v6.b }[2], [x12], #0x1\n"
- "ld1 { v4.b }[2], [x10], #0x1\n"
- "ld1 { v3.b }[2], [x9], #0x1\n"
- "ld1 { v2.b }[2], [x26], #0x1\n"
- "ld1 { v1.b }[2], [x25], #0x1\n"
+ "ld1 { v8.b }[2], [x12], #0x1\n"
+ "ld1 { v26.b }[2], [x10], #0x1\n"
+ "ld1 { v23.b }[2], [x9], #0x1\n"
+ "ld1 { v3.b }[2], [x26], #0x1\n"
+ "ld1 { v10.b }[2], [x21], #0x1\n"
"b 11f\n"
"10:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
- "ldr b9, [x15], #0x1\n"
- "ldr b8, [x14], #0x1\n"
+ "ldr b15, [x15], #0x1\n"
+ "ldr b25, [x14], #0x1\n"
"ldr b7, [x13], #0x1\n"
- "ldr b6, [x12], #0x1\n"
- "ldr b4, [x10], #0x1\n"
- "ldr b3, [x9], #0x1\n"
- "ldr b2, [x26], #0x1\n"
- "ldr b1, [x25], #0x1\n"
+ "ldr b8, [x12], #0x1\n"
+ "ldr b26, [x10], #0x1\n"
+ "ldr b23, [x9], #0x1\n"
+ "ldr b3, [x26], #0x1\n"
+ "ldr b10, [x21], #0x1\n"
"11:" // Oddments: Load (A): Bit 3: End
"ldp x15, x14, [%x[inptrs], #0x40]\n"
"ldp x13, x12, [%x[inptrs], #0x50]\n"
"add x15, x15, x28\n"
"add x14, x14, x28\n"
"ldp x10, x9, [%x[inptrs], #0x60]\n"
- "ldp x26, x25, [%x[inptrs], #0x70]\n"
+ "ldp x26, x21, [%x[inptrs], #0x70]\n"
"add x13, x13, x28\n"
"add x12, x12, x28\n"
"add x10, x10, x28\n"
"add x9, x9, x28\n"
"add x26, x26, x28\n"
- "add x25, x25, x28\n"
+ "add x21, x21, x28\n"
"tbz %x[n_channels], #3, 15f\n"
- "ldr d29, [x15], #0x8\n"
- "ldr d28, [x14], #0x8\n"
- "ldr d27, [x13], #0x8\n"
- "ldr d26, [x12], #0x8\n"
- "ldr d24, [x10], #0x8\n"
- "ldr d23, [x9], #0x8\n"
- "ldr d22, [x26], #0x8\n"
- "ldr d21, [x25], #0x8\n"
+ "ldr d22, [x15], #0x8\n"
+ "ldr d19, [x14], #0x8\n"
+ "ldr d0, [x13], #0x8\n"
+ "ldr d5, [x12], #0x8\n"
+ "ldr d27, [x10], #0x8\n"
+ "ldr d24, [x9], #0x8\n"
+ "ldr d2, [x26], #0x8\n"
+ "ldr d9, [x21], #0x8\n"
"tbz %x[n_channels], #2, 13f\n"
- "ld1 { v29.s }[2], [x15], #0x4\n"
- "ld1 { v28.s }[2], [x14], #0x4\n"
- "ld1 { v27.s }[2], [x13], #0x4\n"
- "ld1 { v26.s }[2], [x12], #0x4\n"
- "ld1 { v24.s }[2], [x10], #0x4\n"
- "ld1 { v23.s }[2], [x9], #0x4\n"
- "ld1 { v22.s }[2], [x26], #0x4\n"
- "ld1 { v21.s }[2], [x25], #0x4\n"
+ "ld1 { v22.s }[2], [x15], #0x4\n"
+ "ld1 { v19.s }[2], [x14], #0x4\n"
+ "ld1 { v0.s }[2], [x13], #0x4\n"
+ "ld1 { v5.s }[2], [x12], #0x4\n"
+ "ld1 { v27.s }[2], [x10], #0x4\n"
+ "ld1 { v24.s }[2], [x9], #0x4\n"
+ "ld1 { v2.s }[2], [x26], #0x4\n"
+ "ld1 { v9.s }[2], [x21], #0x4\n"
"tbz %x[n_channels], #1, 12f\n"
- "ld1 { v29.h }[6], [x15], #0x2\n"
- "ld1 { v28.h }[6], [x14], #0x2\n"
- "ld1 { v27.h }[6], [x13], #0x2\n"
- "ld1 { v26.h }[6], [x12], #0x2\n"
- "ld1 { v24.h }[6], [x10], #0x2\n"
- "ld1 { v23.h }[6], [x9], #0x2\n"
- "ld1 { v22.h }[6], [x26], #0x2\n"
- "ld1 { v21.h }[6], [x25], #0x2\n"
+ "ld1 { v22.h }[6], [x15], #0x2\n"
+ "ld1 { v19.h }[6], [x14], #0x2\n"
+ "ld1 { v0.h }[6], [x13], #0x2\n"
+ "ld1 { v5.h }[6], [x12], #0x2\n"
+ "ld1 { v27.h }[6], [x10], #0x2\n"
+ "ld1 { v24.h }[6], [x9], #0x2\n"
+ "ld1 { v2.h }[6], [x26], #0x2\n"
+ "ld1 { v9.h }[6], [x21], #0x2\n"
"tbz %x[n_channels], #0, 19f\n"
- "ld1 { v29.b }[14], [x15], #0x1\n"
- "ld1 { v28.b }[14], [x14], #0x1\n"
- "ld1 { v27.b }[14], [x13], #0x1\n"
- "ld1 { v26.b }[14], [x12], #0x1\n"
- "ld1 { v24.b }[14], [x10], #0x1\n"
- "ld1 { v23.b }[14], [x9], #0x1\n"
- "ld1 { v22.b }[14], [x26], #0x1\n"
- "ld1 { v21.b }[14], [x25], #0x1\n"
+ "ld1 { v22.b }[14], [x15], #0x1\n"
+ "ld1 { v19.b }[14], [x14], #0x1\n"
+ "ld1 { v0.b }[14], [x13], #0x1\n"
+ "ld1 { v5.b }[14], [x12], #0x1\n"
+ "ld1 { v27.b }[14], [x10], #0x1\n"
+ "ld1 { v24.b }[14], [x9], #0x1\n"
+ "ld1 { v2.b }[14], [x26], #0x1\n"
+ "ld1 { v9.b }[14], [x21], #0x1\n"
"b 19f\n"
"12:" // Oddments: Load (B): Bit 3: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 19f\n"
- "ld1 { v29.b }[12], [x15], #0x1\n"
- "ld1 { v28.b }[12], [x14], #0x1\n"
- "ld1 { v27.b }[12], [x13], #0x1\n"
- "ld1 { v26.b }[12], [x12], #0x1\n"
- "ld1 { v24.b }[12], [x10], #0x1\n"
- "ld1 { v23.b }[12], [x9], #0x1\n"
- "ld1 { v22.b }[12], [x26], #0x1\n"
- "ld1 { v21.b }[12], [x25], #0x1\n"
+ "ld1 { v22.b }[12], [x15], #0x1\n"
+ "ld1 { v19.b }[12], [x14], #0x1\n"
+ "ld1 { v0.b }[12], [x13], #0x1\n"
+ "ld1 { v5.b }[12], [x12], #0x1\n"
+ "ld1 { v27.b }[12], [x10], #0x1\n"
+ "ld1 { v24.b }[12], [x9], #0x1\n"
+ "ld1 { v2.b }[12], [x26], #0x1\n"
+ "ld1 { v9.b }[12], [x21], #0x1\n"
"b 19f\n"
"13:" // Oddments: Load (B): Bit 3: Bit 2: Unset
"tbz %x[n_channels], #1, 14f\n"
- "ld1 { v29.h }[4], [x15], #0x2\n"
- "ld1 { v28.h }[4], [x14], #0x2\n"
- "ld1 { v27.h }[4], [x13], #0x2\n"
- "ld1 { v26.h }[4], [x12], #0x2\n"
- "ld1 { v24.h }[4], [x10], #0x2\n"
- "ld1 { v23.h }[4], [x9], #0x2\n"
- "ld1 { v22.h }[4], [x26], #0x2\n"
- "ld1 { v21.h }[4], [x25], #0x2\n"
+ "ld1 { v22.h }[4], [x15], #0x2\n"
+ "ld1 { v19.h }[4], [x14], #0x2\n"
+ "ld1 { v0.h }[4], [x13], #0x2\n"
+ "ld1 { v5.h }[4], [x12], #0x2\n"
+ "ld1 { v27.h }[4], [x10], #0x2\n"
+ "ld1 { v24.h }[4], [x9], #0x2\n"
+ "ld1 { v2.h }[4], [x26], #0x2\n"
+ "ld1 { v9.h }[4], [x21], #0x2\n"
"tbz %x[n_channels], #0, 19f\n"
- "ld1 { v29.b }[10], [x15], #0x1\n"
- "ld1 { v28.b }[10], [x14], #0x1\n"
- "ld1 { v27.b }[10], [x13], #0x1\n"
- "ld1 { v26.b }[10], [x12], #0x1\n"
- "ld1 { v24.b }[10], [x10], #0x1\n"
- "ld1 { v23.b }[10], [x9], #0x1\n"
- "ld1 { v22.b }[10], [x26], #0x1\n"
- "ld1 { v21.b }[10], [x25], #0x1\n"
+ "ld1 { v22.b }[10], [x15], #0x1\n"
+ "ld1 { v19.b }[10], [x14], #0x1\n"
+ "ld1 { v0.b }[10], [x13], #0x1\n"
+ "ld1 { v5.b }[10], [x12], #0x1\n"
+ "ld1 { v27.b }[10], [x10], #0x1\n"
+ "ld1 { v24.b }[10], [x9], #0x1\n"
+ "ld1 { v2.b }[10], [x26], #0x1\n"
+ "ld1 { v9.b }[10], [x21], #0x1\n"
"b 19f\n"
"14:" // Oddments: Load (B): Bit 3: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 19f\n"
- "ld1 { v29.b }[8], [x15], #0x1\n"
- "ld1 { v28.b }[8], [x14], #0x1\n"
- "ld1 { v27.b }[8], [x13], #0x1\n"
- "ld1 { v26.b }[8], [x12], #0x1\n"
- "ld1 { v24.b }[8], [x10], #0x1\n"
- "ld1 { v23.b }[8], [x9], #0x1\n"
- "ld1 { v22.b }[8], [x26], #0x1\n"
- "ld1 { v21.b }[8], [x25], #0x1\n"
+ "ld1 { v22.b }[8], [x15], #0x1\n"
+ "ld1 { v19.b }[8], [x14], #0x1\n"
+ "ld1 { v0.b }[8], [x13], #0x1\n"
+ "ld1 { v5.b }[8], [x12], #0x1\n"
+ "ld1 { v27.b }[8], [x10], #0x1\n"
+ "ld1 { v24.b }[8], [x9], #0x1\n"
+ "ld1 { v2.b }[8], [x26], #0x1\n"
+ "ld1 { v9.b }[8], [x21], #0x1\n"
"b 19f\n"
"15:" // Oddments: Load (B): Bit 3: Unset
"tbz %x[n_channels], #2, 17f\n"
- "ldr s29, [x15], #0x4\n"
- "ldr s28, [x14], #0x4\n"
- "ldr s27, [x13], #0x4\n"
- "ldr s26, [x12], #0x4\n"
- "ldr s24, [x10], #0x4\n"
- "ldr s23, [x9], #0x4\n"
- "ldr s22, [x26], #0x4\n"
- "ldr s21, [x25], #0x4\n"
+ "ldr s22, [x15], #0x4\n"
+ "ldr s19, [x14], #0x4\n"
+ "ldr s0, [x13], #0x4\n"
+ "ldr s5, [x12], #0x4\n"
+ "ldr s27, [x10], #0x4\n"
+ "ldr s24, [x9], #0x4\n"
+ "ldr s2, [x26], #0x4\n"
+ "ldr s9, [x21], #0x4\n"
"tbz %x[n_channels], #1, 16f\n"
- "ld1 { v29.h }[2], [x15], #0x2\n"
- "ld1 { v28.h }[2], [x14], #0x2\n"
- "ld1 { v27.h }[2], [x13], #0x2\n"
- "ld1 { v26.h }[2], [x12], #0x2\n"
- "ld1 { v24.h }[2], [x10], #0x2\n"
- "ld1 { v23.h }[2], [x9], #0x2\n"
- "ld1 { v22.h }[2], [x26], #0x2\n"
- "ld1 { v21.h }[2], [x25], #0x2\n"
+ "ld1 { v22.h }[2], [x15], #0x2\n"
+ "ld1 { v19.h }[2], [x14], #0x2\n"
+ "ld1 { v0.h }[2], [x13], #0x2\n"
+ "ld1 { v5.h }[2], [x12], #0x2\n"
+ "ld1 { v27.h }[2], [x10], #0x2\n"
+ "ld1 { v24.h }[2], [x9], #0x2\n"
+ "ld1 { v2.h }[2], [x26], #0x2\n"
+ "ld1 { v9.h }[2], [x21], #0x2\n"
"tbz %x[n_channels], #0, 19f\n"
- "ld1 { v29.b }[6], [x15], #0x1\n"
- "ld1 { v28.b }[6], [x14], #0x1\n"
- "ld1 { v27.b }[6], [x13], #0x1\n"
- "ld1 { v26.b }[6], [x12], #0x1\n"
- "ld1 { v24.b }[6], [x10], #0x1\n"
- "ld1 { v23.b }[6], [x9], #0x1\n"
- "ld1 { v22.b }[6], [x26], #0x1\n"
- "ld1 { v21.b }[6], [x25], #0x1\n"
+ "ld1 { v22.b }[6], [x15], #0x1\n"
+ "ld1 { v19.b }[6], [x14], #0x1\n"
+ "ld1 { v0.b }[6], [x13], #0x1\n"
+ "ld1 { v5.b }[6], [x12], #0x1\n"
+ "ld1 { v27.b }[6], [x10], #0x1\n"
+ "ld1 { v24.b }[6], [x9], #0x1\n"
+ "ld1 { v2.b }[6], [x26], #0x1\n"
+ "ld1 { v9.b }[6], [x21], #0x1\n"
"b 19f\n"
"16:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 19f\n"
- "ld1 { v29.b }[4], [x15], #0x1\n"
- "ld1 { v28.b }[4], [x14], #0x1\n"
- "ld1 { v27.b }[4], [x13], #0x1\n"
- "ld1 { v26.b }[4], [x12], #0x1\n"
- "ld1 { v24.b }[4], [x10], #0x1\n"
- "ld1 { v23.b }[4], [x9], #0x1\n"
- "ld1 { v22.b }[4], [x26], #0x1\n"
- "ld1 { v21.b }[4], [x25], #0x1\n"
+ "ld1 { v22.b }[4], [x15], #0x1\n"
+ "ld1 { v19.b }[4], [x14], #0x1\n"
+ "ld1 { v0.b }[4], [x13], #0x1\n"
+ "ld1 { v5.b }[4], [x12], #0x1\n"
+ "ld1 { v27.b }[4], [x10], #0x1\n"
+ "ld1 { v24.b }[4], [x9], #0x1\n"
+ "ld1 { v2.b }[4], [x26], #0x1\n"
+ "ld1 { v9.b }[4], [x21], #0x1\n"
"b 19f\n"
"17:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Unset
"tbz %x[n_channels], #1, 18f\n"
- "ldr h29, [x15], #0x2\n"
- "ldr h28, [x14], #0x2\n"
- "ldr h27, [x13], #0x2\n"
- "ldr h26, [x12], #0x2\n"
- "ldr h24, [x10], #0x2\n"
- "ldr h23, [x9], #0x2\n"
- "ldr h22, [x26], #0x2\n"
- "ldr h21, [x25], #0x2\n"
+ "ldr h22, [x15], #0x2\n"
+ "ldr h19, [x14], #0x2\n"
+ "ldr h0, [x13], #0x2\n"
+ "ldr h5, [x12], #0x2\n"
+ "ldr h27, [x10], #0x2\n"
+ "ldr h24, [x9], #0x2\n"
+ "ldr h2, [x26], #0x2\n"
+ "ldr h9, [x21], #0x2\n"
"tbz %x[n_channels], #0, 19f\n"
- "ld1 { v29.b }[2], [x15], #0x1\n"
- "ld1 { v28.b }[2], [x14], #0x1\n"
- "ld1 { v27.b }[2], [x13], #0x1\n"
- "ld1 { v26.b }[2], [x12], #0x1\n"
- "ld1 { v24.b }[2], [x10], #0x1\n"
- "ld1 { v23.b }[2], [x9], #0x1\n"
- "ld1 { v22.b }[2], [x26], #0x1\n"
- "ld1 { v21.b }[2], [x25], #0x1\n"
+ "ld1 { v22.b }[2], [x15], #0x1\n"
+ "ld1 { v19.b }[2], [x14], #0x1\n"
+ "ld1 { v0.b }[2], [x13], #0x1\n"
+ "ld1 { v5.b }[2], [x12], #0x1\n"
+ "ld1 { v27.b }[2], [x10], #0x1\n"
+ "ld1 { v24.b }[2], [x9], #0x1\n"
+ "ld1 { v2.b }[2], [x26], #0x1\n"
+ "ld1 { v9.b }[2], [x21], #0x1\n"
"b 19f\n"
"18:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
- "ldr b29, [x15], #0x1\n"
- "ldr b28, [x14], #0x1\n"
- "ldr b27, [x13], #0x1\n"
- "ldr b26, [x12], #0x1\n"
- "ldr b24, [x10], #0x1\n"
- "ldr b23, [x9], #0x1\n"
- "ldr b22, [x26], #0x1\n"
- "ldr b21, [x25], #0x1\n"
+ "ldr b22, [x15], #0x1\n"
+ "ldr b19, [x14], #0x1\n"
+ "ldr b0, [x13], #0x1\n"
+ "ldr b5, [x12], #0x1\n"
+ "ldr b27, [x10], #0x1\n"
+ "ldr b24, [x9], #0x1\n"
+ "ldr b2, [x26], #0x1\n"
+ "ldr b9, [x21], #0x1\n"
"19:" // Oddments: Load (B): Bit 3: End
- "ldr q0, [%x[params], #0x10]\n"
- "ldr q16, [%x[params], #0x20]\n"
- "zip2 v30.16b, v4.16b, v2.16b\n"
- "zip1 v4.16b, v4.16b, v2.16b\n"
- "ldr q31, [%x[params], #0x30]\n"
- "zip1 v2.16b, v3.16b, v1.16b\n"
- "zip2 v5.16b, v9.16b, v7.16b\n"
+ "ldr q20, [%x[params], #0x10]\n"
+ "ldr q6, [%x[params], #0x20]\n"
+ "zip2 v1.16b, v26.16b, v3.16b\n"
+ "zip1 v26.16b, v26.16b, v3.16b\n"
+ "ldr q4, [%x[params], #0x30]\n"
+ "zip1 v18.16b, v23.16b, v10.16b\n"
+ "zip2 v30.16b, v15.16b, v7.16b\n"
"cmp x20, #0x4\n"
- "zip1 v9.16b, v9.16b, v7.16b\n"
- "zip1 v7.16b, v8.16b, v6.16b\n"
- "zip2 v6.16b, v8.16b, v6.16b\n"
- "zip2 v1.16b, v3.16b, v1.16b\n"
- "zip2 v3.16b, v4.16b, v2.16b\n"
- "zip1 v4.16b, v4.16b, v2.16b\n"
- "zip2 v25.16b, v29.16b, v27.16b\n"
- "zip1 v29.16b, v29.16b, v27.16b\n"
- "zip1 v27.16b, v28.16b, v26.16b\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8495d3 // udot v19.4s, v14.16b, v4.16b\n"
- "zip2 v8.16b, v9.16b, v7.16b\n"
- "zip1 v9.16b, v9.16b, v7.16b\n"
- "zip1 v7.16b, v5.16b, v6.16b\n"
- "zip2 v6.16b, v5.16b, v6.16b\n"
- "ldr q5, [%x[params], #0x0]\n"
- "zip2 v26.16b, v28.16b, v26.16b\n"
- "zip2 v20.16b, v24.16b, v22.16b\n"
- "zip1 v24.16b, v24.16b, v22.16b\n"
- "zip1 v22.16b, v23.16b, v21.16b\n"
- "zip2 v21.16b, v23.16b, v21.16b\n"
- "zip2 v28.16b, v29.16b, v27.16b\n"
- "zip1 v29.16b, v29.16b, v27.16b\n"
- "zip1 v2.16b, v30.16b, v1.16b\n"
- ".inst 0x6e9d95d3 // udot v19.4s, v14.16b, v29.16b\n"
- "zip2 v1.16b, v30.16b, v1.16b\n"
- "zip1 v27.16b, v25.16b, v26.16b\n"
- "zip2 v26.16b, v25.16b, v26.16b\n"
- "zip2 v23.16b, v24.16b, v22.16b\n"
- "zip1 v24.16b, v24.16b, v22.16b\n"
- "zip1 v22.16b, v20.16b, v21.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9895d2 // udot v18.4s, v14.16b, v24.16b\n"
- "zip2 v21.16b, v20.16b, v21.16b\n"
- "mov v30.16b, v5.16b\n"
- ".inst 0x6e8995d3 // udot v19.4s, v14.16b, v9.16b\n"
- "mov v25.16b, v5.16b\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e899405 // udot v5.4s, v0.16b, v9.16b\n"
- ".inst 0x6e849419 // udot v25.4s, v0.16b, v4.16b\n"
- ".inst 0x6e849605 // udot v5.4s, v16.16b, v4.16b\n"
- "ext v4.16b, v4.16b, v4.16b, #0x1\n"
- "ext v9.16b, v9.16b, v9.16b, #0x1\n"
- ".inst 0x6e9d9619 // udot v25.4s, v16.16b, v29.16b\n"
- ".inst 0x6e9d97e5 // udot v5.4s, v31.16b, v29.16b\n"
- "ext v29.16b, v29.16b, v29.16b, #0x1\n"
- ".inst 0x6e89941e // udot v30.4s, v0.16b, v9.16b\n"
- ".inst 0x6e849414 // udot v20.4s, v0.16b, v4.16b\n"
+ "zip1 v15.16b, v15.16b, v7.16b\n"
+ "zip1 v29.16b, v25.16b, v8.16b\n"
+ "zip2 v8.16b, v25.16b, v8.16b\n"
+ "zip2 v10.16b, v23.16b, v10.16b\n"
+ "zip2 v23.16b, v26.16b, v18.16b\n"
+ "zip1 v26.16b, v26.16b, v18.16b\n"
+ "zip2 v28.16b, v22.16b, v0.16b\n"
+ "zip1 v22.16b, v22.16b, v0.16b\n"
+ "zip1 v21.16b, v19.16b, v5.16b\n"
"movi v17.4s, #0x0\n"
- ".inst 0x6e8495d1 // udot v17.4s, v14.16b, v4.16b\n"
- ".inst 0x6e9d95d1 // udot v17.4s, v14.16b, v29.16b\n"
- ".inst 0x6e9897f9 // udot v25.4s, v31.16b, v24.16b\n"
- "ext v24.16b, v24.16b, v24.16b, #0x1\n"
- ".inst 0x6e84961e // udot v30.4s, v16.16b, v4.16b\n"
- "ldr q4, [%x[params], #0x50]\n"
- ".inst 0x6e9d9614 // udot v20.4s, v16.16b, v29.16b\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9895d0 // udot v16.4s, v14.16b, v24.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e8995d1 // udot v17.4s, v14.16b, v9.16b\n"
- "ldr q9, [%x[params], #0x40]\n"
- ".inst 0x6e9d97fe // udot v30.4s, v31.16b, v29.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- ".inst 0x6e9897f4 // udot v20.4s, v31.16b, v24.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
+ ".inst 0x6e9a9591 // udot v17.4s, v12.16b, v26.16b\n"
+ "zip2 v25.16b, v15.16b, v29.16b\n"
+ "zip1 v15.16b, v15.16b, v29.16b\n"
+ "zip1 v7.16b, v30.16b, v8.16b\n"
+ "zip2 v8.16b, v30.16b, v8.16b\n"
+ "ldr q31, [%x[params], #0x0]\n"
+ "zip2 v5.16b, v19.16b, v5.16b\n"
+ "zip2 v30.16b, v27.16b, v2.16b\n"
+ "zip1 v27.16b, v27.16b, v2.16b\n"
+ "zip1 v18.16b, v24.16b, v9.16b\n"
+ "zip2 v9.16b, v24.16b, v9.16b\n"
+ "zip2 v19.16b, v22.16b, v21.16b\n"
+ "zip1 v22.16b, v22.16b, v21.16b\n"
+ "zip1 v3.16b, v1.16b, v10.16b\n"
+ ".inst 0x6e969591 // udot v17.4s, v12.16b, v22.16b\n"
+ "zip2 v10.16b, v1.16b, v10.16b\n"
+ "zip1 v0.16b, v28.16b, v5.16b\n"
+ "zip2 v5.16b, v28.16b, v5.16b\n"
+ "zip2 v24.16b, v27.16b, v18.16b\n"
+ "zip1 v27.16b, v27.16b, v18.16b\n"
+ "zip1 v2.16b, v30.16b, v9.16b\n"
+ "mov v18.16b, v17.16b\n .inst 0x6e9b9592 // udot v18.4s, v12.16b, v27.16b\n"
+ "zip2 v9.16b, v30.16b, v9.16b\n"
+ "mov v30.16b, v31.16b\n"
+ ".inst 0x6e8f9591 // udot v17.4s, v12.16b, v15.16b\n"
+ "mov v29.16b, v31.16b\n"
+ "mov v28.16b, v31.16b\n"
+ ".inst 0x6e8f969f // udot v31.4s, v20.16b, v15.16b\n"
+ ".inst 0x6e9a969d // udot v29.4s, v20.16b, v26.16b\n"
+ ".inst 0x6e9a94df // udot v31.4s, v6.16b, v26.16b\n"
+ "ext v26.16b, v26.16b, v26.16b, #0x1\n"
+ "movi v1.4s, #0x0\n"
+ "ext v15.16b, v15.16b, v15.16b, #0x1\n"
+ ".inst 0x6e9a9581 // udot v1.4s, v12.16b, v26.16b\n"
+ ".inst 0x6e9694dd // udot v29.4s, v6.16b, v22.16b\n"
+ ".inst 0x6e96949f // udot v31.4s, v4.16b, v22.16b\n"
+ "ext v22.16b, v22.16b, v22.16b, #0x1\n"
+ ".inst 0x6e8f969e // udot v30.4s, v20.16b, v15.16b\n"
+ ".inst 0x6e9a969c // udot v28.4s, v20.16b, v26.16b\n"
+ "mls v31.4s, v17.4s, v16.4s\n"
+ ".inst 0x6e969581 // udot v1.4s, v12.16b, v22.16b\n"
+ ".inst 0x6e9b949d // udot v29.4s, v4.16b, v27.16b\n"
+ "ext v27.16b, v27.16b, v27.16b, #0x1\n"
+ ".inst 0x6e9a94de // udot v30.4s, v6.16b, v26.16b\n"
+ "ldr q21, [%x[params], #0x50]\n"
+ ".inst 0x6e9694dc // udot v28.4s, v6.16b, v22.16b\n"
+ "mls v29.4s, v18.4s, v16.4s\n"
+ "mov v20.16b, v1.16b\n .inst 0x6e9b9594 // udot v20.4s, v12.16b, v27.16b\n"
+ ".inst 0x6e8f9581 // udot v1.4s, v12.16b, v15.16b\n"
+ "ldr q18, [%x[params], #0x40]\n"
+ "sqrdmulh v31.4s, v31.4s, v18.4s\n"
+ ".inst 0x6e96949e // udot v30.4s, v4.16b, v22.16b\n"
+ ".inst 0x6e9b949c // udot v28.4s, v4.16b, v27.16b\n"
+ "mls v30.4s, v1.4s, v16.4s\n"
"add %x[params], %x[params], #0x60\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smax v5.4s, v5.4s, v13.4s\n"
+ "mls v28.4s, v20.4s, v16.4s\n"
+ "and v17.16b, v31.16b, v21.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v18.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v18.4s\n"
+ "sqrdmulh v28.4s, v28.4s, v18.4s\n"
+ "sqadd v31.4s, v31.4s, v17.4s\n"
+ "and v17.16b, v30.16b, v21.16b\n"
+ "and v18.16b, v29.16b, v21.16b\n"
+ "and v26.16b, v28.16b, v21.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v26.4s, v26.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v17.4s\n"
+ "sqadd v29.4s, v29.4s, v18.4s\n"
+ "sqadd v28.4s, v28.4s, v26.4s\n"
+ "srshl v31.4s, v31.4s, v21.4s\n"
+ "srshl v30.4s, v30.4s, v21.4s\n"
+ "srshl v29.4s, v29.4s, v21.4s\n"
+ "srshl v28.4s, v28.4s, v21.4s\n"
+ "add v31.4s, v31.4s, v14.4s\n"
+ "add v30.4s, v30.4s, v14.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v28.4s, v28.4s, v14.4s\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
"smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
+ "smin v30.4s, v30.4s, v11.4s\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v28.4s, v28.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
"blt 20f\n"
- "str s5, [x24, x27]\n"
- "str s30, [x23, x27]\n"
- "str s25, [x22, x27]\n"
- "str s20, [x21, x27]\n"
+ "str s31, [x25, x27]\n"
+ "str s30, [x24, x27]\n"
+ "str s29, [x23, x27]\n"
+ "str s28, [x22, x27]\n"
"b 23f\n"
"20:" // Oddments: Unroll 0: Oddment store
+ "add x25, x25, x27\n"
"add x24, x24, x27\n"
"add x23, x23, x27\n"
"add x22, x22, x27\n"
- "add x21, x21, x27\n"
"tbz x20, #1, 21f\n"
- "st1 { v5.h }[0], [x24], #0x2\n"
- "st1 { v30.h }[0], [x23], #0x2\n"
- "st1 { v25.h }[0], [x22], #0x2\n"
- "st1 { v20.h }[0], [x21], #0x2\n"
+ "st1 { v31.h }[0], [x25], #0x2\n"
+ "st1 { v30.h }[0], [x24], #0x2\n"
+ "st1 { v29.h }[0], [x23], #0x2\n"
+ "st1 { v28.h }[0], [x22], #0x2\n"
"tbz x20, #0, 22f\n"
- "st1 { v5.b }[2], [x24], #0x1\n"
- "st1 { v30.b }[2], [x23], #0x1\n"
- "st1 { v25.b }[2], [x22], #0x1\n"
- "st1 { v20.b }[2], [x21], #0x1\n"
+ "st1 { v31.b }[2], [x25], #0x1\n"
+ "st1 { v30.b }[2], [x24], #0x1\n"
+ "st1 { v29.b }[2], [x23], #0x1\n"
+ "st1 { v28.b }[2], [x22], #0x1\n"
"b 22f\n"
"21:" // Oddments: Unroll 0: Oddment store: Bit 1: Unset
- "st1 { v5.b }[0], [x24], #0x1\n"
- "st1 { v30.b }[0], [x23], #0x1\n"
- "st1 { v25.b }[0], [x22], #0x1\n"
- "st1 { v20.b }[0], [x21], #0x1\n"
+ "st1 { v31.b }[0], [x25], #0x1\n"
+ "st1 { v30.b }[0], [x24], #0x1\n"
+ "st1 { v29.b }[0], [x23], #0x1\n"
+ "st1 { v28.b }[0], [x22], #0x1\n"
"22:" // Oddments: Unroll 0: Oddment store: Bit 1: End
"23:" // Oddments: Unroll 0: After oddment store
"subs x20, x20, #0x4\n"
"add x27, x27, #0x4\n"
"ble 35f\n"
- "ldr q5, [%x[params], #0x0]\n"
- "ldr q0, [%x[params], #0x10]\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8395d3 // udot v19.4s, v14.16b, v3.16b\n"
- "ldr q16, [%x[params], #0x20]\n"
- "ldr q31, [%x[params], #0x30]\n"
- "mov v30.16b, v5.16b\n"
- "mov v25.16b, v5.16b\n"
- "ldr q9, [%x[params], #0x40]\n"
- "ldr q4, [%x[params], #0x50]\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e889405 // udot v5.4s, v0.16b, v8.16b\n"
- ".inst 0x6e9c95d3 // udot v19.4s, v14.16b, v28.16b\n"
- ".inst 0x6e839419 // udot v25.4s, v0.16b, v3.16b\n"
- "movi v17.4s, #0x0\n"
+ "ldr q31, [%x[params], #0x0]\n"
+ "ldr q27, [%x[params], #0x10]\n"
+ "movi v1.4s, #0x0\n"
+ ".inst 0x6e979581 // udot v1.4s, v12.16b, v23.16b\n"
+ "ldr q26, [%x[params], #0x20]\n"
+ "ldr q22, [%x[params], #0x30]\n"
+ "mov v30.16b, v31.16b\n"
+ "mov v29.16b, v31.16b\n"
+ "ldr q4, [%x[params], #0x40]\n"
+ "ldr q21, [%x[params], #0x50]\n"
+ "mov v28.16b, v31.16b\n"
+ ".inst 0x6e99977f // udot v31.4s, v27.16b, v25.16b\n"
+ ".inst 0x6e939581 // udot v1.4s, v12.16b, v19.16b\n"
+ ".inst 0x6e97977d // udot v29.4s, v27.16b, v23.16b\n"
+ "movi v20.4s, #0x0\n"
"cmp x20, #0x4\n"
- ".inst 0x6e839605 // udot v5.4s, v16.16b, v3.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9795d2 // udot v18.4s, v14.16b, v23.16b\n"
- "ext v3.16b, v3.16b, v3.16b, #0x1\n"
- "add %x[params], %x[params], #0x60\n"
- ".inst 0x6e8895d3 // udot v19.4s, v14.16b, v8.16b\n"
- "ext v8.16b, v8.16b, v8.16b, #0x1\n"
- ".inst 0x6e88941e // udot v30.4s, v0.16b, v8.16b\n"
- ".inst 0x6e839414 // udot v20.4s, v0.16b, v3.16b\n"
- ".inst 0x6e8395d1 // udot v17.4s, v14.16b, v3.16b\n"
- ".inst 0x6e9c9619 // udot v25.4s, v16.16b, v28.16b\n"
- ".inst 0x6e9c97e5 // udot v5.4s, v31.16b, v28.16b\n"
- "ext v28.16b, v28.16b, v28.16b, #0x1\n"
- ".inst 0x6e83961e // udot v30.4s, v16.16b, v3.16b\n"
- ".inst 0x6e9c9614 // udot v20.4s, v16.16b, v28.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9c95d1 // udot v17.4s, v14.16b, v28.16b\n"
- ".inst 0x6e9797f9 // udot v25.4s, v31.16b, v23.16b\n"
+ ".inst 0x6e97975f // udot v31.4s, v26.16b, v23.16b\n"
+ "mov v18.16b, v1.16b\n .inst 0x6e989592 // udot v18.4s, v12.16b, v24.16b\n"
"ext v23.16b, v23.16b, v23.16b, #0x1\n"
- ".inst 0x6e9c97fe // udot v30.4s, v31.16b, v28.16b\n"
- ".inst 0x6e9797f4 // udot v20.4s, v31.16b, v23.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9795d0 // udot v16.4s, v14.16b, v23.16b\n"
- ".inst 0x6e8895d1 // udot v17.4s, v14.16b, v8.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smax v5.4s, v5.4s, v13.4s\n"
+ "add %x[params], %x[params], #0x60\n"
+ ".inst 0x6e999581 // udot v1.4s, v12.16b, v25.16b\n"
+ "ext v25.16b, v25.16b, v25.16b, #0x1\n"
+ ".inst 0x6e99977e // udot v30.4s, v27.16b, v25.16b\n"
+ ".inst 0x6e97977c // udot v28.4s, v27.16b, v23.16b\n"
+ ".inst 0x6e979594 // udot v20.4s, v12.16b, v23.16b\n"
+ ".inst 0x6e93975d // udot v29.4s, v26.16b, v19.16b\n"
+ ".inst 0x6e9396df // udot v31.4s, v22.16b, v19.16b\n"
+ "ext v19.16b, v19.16b, v19.16b, #0x1\n"
+ ".inst 0x6e97975e // udot v30.4s, v26.16b, v23.16b\n"
+ ".inst 0x6e93975c // udot v28.4s, v26.16b, v19.16b\n"
+ "mls v31.4s, v1.4s, v16.4s\n"
+ ".inst 0x6e939594 // udot v20.4s, v12.16b, v19.16b\n"
+ ".inst 0x6e9896dd // udot v29.4s, v22.16b, v24.16b\n"
+ "ext v24.16b, v24.16b, v24.16b, #0x1\n"
+ ".inst 0x6e9396de // udot v30.4s, v22.16b, v19.16b\n"
+ ".inst 0x6e9896dc // udot v28.4s, v22.16b, v24.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v4.4s\n"
+ "mov v17.16b, v20.16b\n .inst 0x6e989591 // udot v17.4s, v12.16b, v24.16b\n"
+ ".inst 0x6e999594 // udot v20.4s, v12.16b, v25.16b\n"
+ "mls v30.4s, v20.4s, v16.4s\n"
+ "mls v29.4s, v18.4s, v16.4s\n"
+ "mls v28.4s, v17.4s, v16.4s\n"
+ "and v17.16b, v31.16b, v21.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v4.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v4.4s\n"
+ "sqrdmulh v28.4s, v28.4s, v4.4s\n"
+ "sqadd v31.4s, v31.4s, v17.4s\n"
+ "and v19.16b, v30.16b, v21.16b\n"
+ "and v18.16b, v29.16b, v21.16b\n"
+ "and v17.16b, v28.16b, 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 v30.4s, v30.4s, v19.4s\n"
+ "sqadd v29.4s, v29.4s, v18.4s\n"
+ "sqadd v28.4s, v28.4s, v17.4s\n"
+ "srshl v31.4s, v31.4s, v21.4s\n"
+ "srshl v30.4s, v30.4s, v21.4s\n"
+ "srshl v29.4s, v29.4s, v21.4s\n"
+ "srshl v28.4s, v28.4s, v21.4s\n"
+ "add v31.4s, v31.4s, v14.4s\n"
+ "add v30.4s, v30.4s, v14.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v28.4s, v28.4s, v14.4s\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
"smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
+ "smin v30.4s, v30.4s, v11.4s\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v28.4s, v28.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
"blt 24f\n"
- "str s5, [x24, x27]\n"
- "str s30, [x23, x27]\n"
- "str s25, [x22, x27]\n"
- "str s20, [x21, x27]\n"
+ "str s31, [x25, x27]\n"
+ "str s30, [x24, x27]\n"
+ "str s29, [x23, x27]\n"
+ "str s28, [x22, x27]\n"
"b 27f\n"
"24:" // Oddments: Unroll 1: Oddment store
+ "add x25, x25, x27\n"
"add x24, x24, x27\n"
"add x23, x23, x27\n"
"add x22, x22, x27\n"
- "add x21, x21, x27\n"
"tbz x20, #1, 25f\n"
- "st1 { v5.h }[0], [x24], #0x2\n"
- "st1 { v30.h }[0], [x23], #0x2\n"
- "st1 { v25.h }[0], [x22], #0x2\n"
- "st1 { v20.h }[0], [x21], #0x2\n"
+ "st1 { v31.h }[0], [x25], #0x2\n"
+ "st1 { v30.h }[0], [x24], #0x2\n"
+ "st1 { v29.h }[0], [x23], #0x2\n"
+ "st1 { v28.h }[0], [x22], #0x2\n"
"tbz x20, #0, 26f\n"
- "st1 { v5.b }[2], [x24], #0x1\n"
- "st1 { v30.b }[2], [x23], #0x1\n"
- "st1 { v25.b }[2], [x22], #0x1\n"
- "st1 { v20.b }[2], [x21], #0x1\n"
+ "st1 { v31.b }[2], [x25], #0x1\n"
+ "st1 { v30.b }[2], [x24], #0x1\n"
+ "st1 { v29.b }[2], [x23], #0x1\n"
+ "st1 { v28.b }[2], [x22], #0x1\n"
"b 26f\n"
"25:" // Oddments: Unroll 1: Oddment store: Bit 1: Unset
- "st1 { v5.b }[0], [x24], #0x1\n"
- "st1 { v30.b }[0], [x23], #0x1\n"
- "st1 { v25.b }[0], [x22], #0x1\n"
- "st1 { v20.b }[0], [x21], #0x1\n"
+ "st1 { v31.b }[0], [x25], #0x1\n"
+ "st1 { v30.b }[0], [x24], #0x1\n"
+ "st1 { v29.b }[0], [x23], #0x1\n"
+ "st1 { v28.b }[0], [x22], #0x1\n"
"26:" // Oddments: Unroll 1: Oddment store: Bit 1: End
"27:" // Oddments: Unroll 1: After oddment store
"subs x20, x20, #0x4\n"
"add x27, x27, #0x4\n"
"ble 35f\n"
- "ldr q5, [%x[params], #0x0]\n"
- "ldr q0, [%x[params], #0x10]\n"
+ "ldr q31, [%x[params], #0x0]\n"
+ "ldr q25, [%x[params], #0x10]\n"
+ "movi v24.4s, #0x0\n"
+ ".inst 0x6e839598 // udot v24.4s, v12.16b, v3.16b\n"
+ "ldr q23, [%x[params], #0x20]\n"
+ "ldr q22, [%x[params], #0x30]\n"
+ "mov v30.16b, v31.16b\n"
+ "mov v29.16b, v31.16b\n"
+ "ldr q21, [%x[params], #0x40]\n"
+ "ldr q20, [%x[params], #0x50]\n"
+ "mov v28.16b, v31.16b\n"
+ ".inst 0x6e87973f // udot v31.4s, v25.16b, v7.16b\n"
+ ".inst 0x6e809598 // udot v24.4s, v12.16b, v0.16b\n"
+ ".inst 0x6e83973d // udot v29.4s, v25.16b, v3.16b\n"
"movi v19.4s, #0x0\n"
- ".inst 0x6e8295d3 // udot v19.4s, v14.16b, v2.16b\n"
- "ldr q16, [%x[params], #0x20]\n"
- "ldr q31, [%x[params], #0x30]\n"
- "mov v30.16b, v5.16b\n"
- "mov v25.16b, v5.16b\n"
- "ldr q9, [%x[params], #0x40]\n"
- "ldr q4, [%x[params], #0x50]\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e879405 // udot v5.4s, v0.16b, v7.16b\n"
- ".inst 0x6e9b95d3 // udot v19.4s, v14.16b, v27.16b\n"
- ".inst 0x6e829419 // udot v25.4s, v0.16b, v2.16b\n"
- "movi v17.4s, #0x0\n"
"cmp x20, #0x4\n"
- ".inst 0x6e829605 // udot v5.4s, v16.16b, v2.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9695d2 // udot v18.4s, v14.16b, v22.16b\n"
- "ext v2.16b, v2.16b, v2.16b, #0x1\n"
+ ".inst 0x6e8396ff // udot v31.4s, v23.16b, v3.16b\n"
+ "mov v18.16b, v24.16b\n .inst 0x6e829592 // udot v18.4s, v12.16b, v2.16b\n"
+ "ext v3.16b, v3.16b, v3.16b, #0x1\n"
"add %x[params], %x[params], #0x60\n"
- ".inst 0x6e8795d3 // udot v19.4s, v14.16b, v7.16b\n"
+ ".inst 0x6e879598 // udot v24.4s, v12.16b, v7.16b\n"
"ext v7.16b, v7.16b, v7.16b, #0x1\n"
- ".inst 0x6e87941e // udot v30.4s, v0.16b, v7.16b\n"
- ".inst 0x6e829414 // udot v20.4s, v0.16b, v2.16b\n"
- ".inst 0x6e8295d1 // udot v17.4s, v14.16b, v2.16b\n"
- ".inst 0x6e9b9619 // udot v25.4s, v16.16b, v27.16b\n"
- ".inst 0x6e9b97e5 // udot v5.4s, v31.16b, v27.16b\n"
- "ext v27.16b, v27.16b, v27.16b, #0x1\n"
- ".inst 0x6e82961e // udot v30.4s, v16.16b, v2.16b\n"
- ".inst 0x6e9b9614 // udot v20.4s, v16.16b, v27.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9b95d1 // udot v17.4s, v14.16b, v27.16b\n"
- ".inst 0x6e9697f9 // udot v25.4s, v31.16b, v22.16b\n"
- "ext v22.16b, v22.16b, v22.16b, #0x1\n"
- ".inst 0x6e9b97fe // udot v30.4s, v31.16b, v27.16b\n"
- ".inst 0x6e9697f4 // udot v20.4s, v31.16b, v22.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9695d0 // udot v16.4s, v14.16b, v22.16b\n"
- ".inst 0x6e8795d1 // udot v17.4s, v14.16b, v7.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smax v5.4s, v5.4s, v13.4s\n"
+ ".inst 0x6e87973e // udot v30.4s, v25.16b, v7.16b\n"
+ ".inst 0x6e83973c // udot v28.4s, v25.16b, v3.16b\n"
+ ".inst 0x6e839593 // udot v19.4s, v12.16b, v3.16b\n"
+ ".inst 0x6e8096fd // udot v29.4s, v23.16b, v0.16b\n"
+ ".inst 0x6e8096df // udot v31.4s, v22.16b, v0.16b\n"
+ "ext v0.16b, v0.16b, v0.16b, #0x1\n"
+ ".inst 0x6e8396fe // udot v30.4s, v23.16b, v3.16b\n"
+ ".inst 0x6e8096fc // udot v28.4s, v23.16b, v0.16b\n"
+ "mls v31.4s, v24.4s, v16.4s\n"
+ ".inst 0x6e809593 // udot v19.4s, v12.16b, v0.16b\n"
+ ".inst 0x6e8296dd // udot v29.4s, v22.16b, v2.16b\n"
+ "ext v2.16b, v2.16b, v2.16b, #0x1\n"
+ ".inst 0x6e8096de // udot v30.4s, v22.16b, v0.16b\n"
+ ".inst 0x6e8296dc // udot v28.4s, v22.16b, v2.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v21.4s\n"
+ "mov v17.16b, v19.16b\n .inst 0x6e829591 // udot v17.4s, v12.16b, v2.16b\n"
+ ".inst 0x6e879593 // udot v19.4s, v12.16b, v7.16b\n"
+ "mls v30.4s, v19.4s, v16.4s\n"
+ "mls v29.4s, v18.4s, v16.4s\n"
+ "mls v28.4s, v17.4s, v16.4s\n"
+ "and v17.16b, v31.16b, v20.16b\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v21.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v21.4s\n"
+ "sqrdmulh v28.4s, v28.4s, v21.4s\n"
+ "sqadd v31.4s, v31.4s, v17.4s\n"
+ "and v19.16b, v30.16b, v20.16b\n"
+ "and v18.16b, v29.16b, v20.16b\n"
+ "and v17.16b, v28.16b, v20.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v19.4s\n"
+ "sqadd v29.4s, v29.4s, v18.4s\n"
+ "sqadd v28.4s, v28.4s, v17.4s\n"
+ "srshl v31.4s, v31.4s, v20.4s\n"
+ "srshl v30.4s, v30.4s, v20.4s\n"
+ "srshl v29.4s, v29.4s, v20.4s\n"
+ "srshl v28.4s, v28.4s, v20.4s\n"
+ "add v31.4s, v31.4s, v14.4s\n"
+ "add v30.4s, v30.4s, v14.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v28.4s, v28.4s, v14.4s\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
"smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
+ "smin v30.4s, v30.4s, v11.4s\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v28.4s, v28.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
"blt 28f\n"
- "str s5, [x24, x27]\n"
- "str s30, [x23, x27]\n"
- "str s25, [x22, x27]\n"
- "str s20, [x21, x27]\n"
+ "str s31, [x25, x27]\n"
+ "str s30, [x24, x27]\n"
+ "str s29, [x23, x27]\n"
+ "str s28, [x22, x27]\n"
"b 31f\n"
"28:" // Oddments: Unroll 2: Oddment store
+ "add x25, x25, x27\n"
"add x24, x24, x27\n"
"add x23, x23, x27\n"
"add x22, x22, x27\n"
- "add x21, x21, x27\n"
"tbz x20, #1, 29f\n"
- "st1 { v5.h }[0], [x24], #0x2\n"
- "st1 { v30.h }[0], [x23], #0x2\n"
- "st1 { v25.h }[0], [x22], #0x2\n"
- "st1 { v20.h }[0], [x21], #0x2\n"
+ "st1 { v31.h }[0], [x25], #0x2\n"
+ "st1 { v30.h }[0], [x24], #0x2\n"
+ "st1 { v29.h }[0], [x23], #0x2\n"
+ "st1 { v28.h }[0], [x22], #0x2\n"
"tbz x20, #0, 30f\n"
- "st1 { v5.b }[2], [x24], #0x1\n"
- "st1 { v30.b }[2], [x23], #0x1\n"
- "st1 { v25.b }[2], [x22], #0x1\n"
- "st1 { v20.b }[2], [x21], #0x1\n"
+ "st1 { v31.b }[2], [x25], #0x1\n"
+ "st1 { v30.b }[2], [x24], #0x1\n"
+ "st1 { v29.b }[2], [x23], #0x1\n"
+ "st1 { v28.b }[2], [x22], #0x1\n"
"b 30f\n"
"29:" // Oddments: Unroll 2: Oddment store: Bit 1: Unset
- "st1 { v5.b }[0], [x24], #0x1\n"
- "st1 { v30.b }[0], [x23], #0x1\n"
- "st1 { v25.b }[0], [x22], #0x1\n"
- "st1 { v20.b }[0], [x21], #0x1\n"
+ "st1 { v31.b }[0], [x25], #0x1\n"
+ "st1 { v30.b }[0], [x24], #0x1\n"
+ "st1 { v29.b }[0], [x23], #0x1\n"
+ "st1 { v28.b }[0], [x22], #0x1\n"
"30:" // Oddments: Unroll 2: Oddment store: Bit 1: End
"31:" // Oddments: Unroll 2: After oddment store
"subs x20, x20, #0x4\n"
"add x27, x27, #0x4\n"
"ble 35f\n"
- "ldr q5, [%x[params], #0x0]\n"
- "ldr q0, [%x[params], #0x10]\n"
- "movi v19.4s, #0x0\n"
- ".inst 0x6e8195d3 // udot v19.4s, v14.16b, v1.16b\n"
- "ldr q16, [%x[params], #0x20]\n"
- "ldr q31, [%x[params], #0x30]\n"
- "mov v30.16b, v5.16b\n"
- "mov v25.16b, v5.16b\n"
- "ldr q9, [%x[params], #0x40]\n"
- "ldr q4, [%x[params], #0x50]\n"
- "mov v20.16b, v5.16b\n"
- ".inst 0x6e869405 // udot v5.4s, v0.16b, v6.16b\n"
- ".inst 0x6e9a95d3 // udot v19.4s, v14.16b, v26.16b\n"
- ".inst 0x6e819419 // udot v25.4s, v0.16b, v1.16b\n"
- "movi v17.4s, #0x0\n"
+ "ldr q31, [%x[params], #0x0]\n"
+ "ldr q23, [%x[params], #0x10]\n"
+ "movi v22.4s, #0x0\n"
+ ".inst 0x6e8a9596 // udot v22.4s, v12.16b, v10.16b\n"
+ "ldr q21, [%x[params], #0x20]\n"
+ "ldr q19, [%x[params], #0x30]\n"
+ "mov v30.16b, v31.16b\n"
+ "mov v29.16b, v31.16b\n"
+ "ldr q20, [%x[params], #0x40]\n"
+ "ldr q26, [%x[params], #0x50]\n"
+ "mov v28.16b, v31.16b\n"
+ ".inst 0x6e8896ff // udot v31.4s, v23.16b, v8.16b\n"
+ ".inst 0x6e859596 // udot v22.4s, v12.16b, v5.16b\n"
+ ".inst 0x6e8a96fd // udot v29.4s, v23.16b, v10.16b\n"
+ "movi v18.4s, #0x0\n"
"add %x[params], %x[params], #0x60\n"
- ".inst 0x6e819605 // udot v5.4s, v16.16b, v1.16b\n"
- "mov v18.16b, v19.16b\n .inst 0x6e9595d2 // udot v18.4s, v14.16b, v21.16b\n"
- "ext v1.16b, v1.16b, v1.16b, #0x1\n"
- ".inst 0x6e8695d3 // udot v19.4s, v14.16b, v6.16b\n"
- "ext v6.16b, v6.16b, v6.16b, #0x1\n"
- ".inst 0x6e86941e // udot v30.4s, v0.16b, v6.16b\n"
- ".inst 0x6e819414 // udot v20.4s, v0.16b, v1.16b\n"
- ".inst 0x6e8195d1 // udot v17.4s, v14.16b, v1.16b\n"
- ".inst 0x6e9a9619 // udot v25.4s, v16.16b, v26.16b\n"
- ".inst 0x6e9a97e5 // udot v5.4s, v31.16b, v26.16b\n"
- "ext v26.16b, v26.16b, v26.16b, #0x1\n"
- ".inst 0x6e81961e // udot v30.4s, v16.16b, v1.16b\n"
- ".inst 0x6e9a9614 // udot v20.4s, v16.16b, v26.16b\n"
- "mls v5.4s, v19.4s, v11.4s\n"
- ".inst 0x6e9a95d1 // udot v17.4s, v14.16b, v26.16b\n"
- ".inst 0x6e9597f9 // udot v25.4s, v31.16b, v21.16b\n"
- "ext v21.16b, v21.16b, v21.16b, #0x1\n"
- ".inst 0x6e9a97fe // udot v30.4s, v31.16b, v26.16b\n"
- ".inst 0x6e9597f4 // udot v20.4s, v31.16b, v21.16b\n"
- "sqrdmulh v5.4s, v5.4s, v9.4s\n"
- "mov v16.16b, v17.16b\n .inst 0x6e9595d0 // udot v16.4s, v14.16b, v21.16b\n"
- ".inst 0x6e8695d1 // udot v17.4s, v14.16b, v6.16b\n"
- "mls v30.4s, v17.4s, v11.4s\n"
- "mls v25.4s, v18.4s, v11.4s\n"
- "mls v20.4s, v16.4s, v11.4s\n"
- "and v0.16b, v5.16b, v4.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqrdmulh v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v25.4s, v25.4s, v9.4s\n"
- "sqrdmulh v20.4s, v20.4s, v9.4s\n"
- "sqadd v5.4s, v5.4s, v0.4s\n"
- "and v16.16b, v30.16b, v4.16b\n"
- "and v31.16b, v25.16b, v4.16b\n"
- "and v0.16b, v20.16b, v4.16b\n"
+ ".inst 0x6e8a96bf // udot v31.4s, v21.16b, v10.16b\n"
+ "mov v17.16b, v22.16b\n .inst 0x6e899591 // udot v17.4s, v12.16b, v9.16b\n"
+ "ext v10.16b, v10.16b, v10.16b, #0x1\n"
+ ".inst 0x6e889596 // udot v22.4s, v12.16b, v8.16b\n"
+ "ext v8.16b, v8.16b, v8.16b, #0x1\n"
+ ".inst 0x6e8896fe // udot v30.4s, v23.16b, v8.16b\n"
+ ".inst 0x6e8a96fc // udot v28.4s, v23.16b, v10.16b\n"
+ ".inst 0x6e8a9592 // udot v18.4s, v12.16b, v10.16b\n"
+ ".inst 0x6e8596bd // udot v29.4s, v21.16b, v5.16b\n"
+ ".inst 0x6e85967f // udot v31.4s, v19.16b, v5.16b\n"
+ "ext v5.16b, v5.16b, v5.16b, #0x1\n"
+ ".inst 0x6e8a96be // udot v30.4s, v21.16b, v10.16b\n"
+ ".inst 0x6e8596bc // udot v28.4s, v21.16b, v5.16b\n"
+ "mls v31.4s, v22.4s, v16.4s\n"
+ ".inst 0x6e859592 // udot v18.4s, v12.16b, v5.16b\n"
+ ".inst 0x6e89967d // udot v29.4s, v19.16b, v9.16b\n"
+ "ext v9.16b, v9.16b, v9.16b, #0x1\n"
+ ".inst 0x6e85967e // udot v30.4s, v19.16b, v5.16b\n"
+ ".inst 0x6e89967c // udot v28.4s, v19.16b, v9.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v20.4s\n"
+ "mov v7.16b, v18.16b\n .inst 0x6e899587 // udot v7.4s, v12.16b, v9.16b\n"
+ ".inst 0x6e889592 // udot v18.4s, v12.16b, v8.16b\n"
+ "mls v30.4s, v18.4s, v16.4s\n"
+ "mls v29.4s, v17.4s, v16.4s\n"
+ "mls v28.4s, v7.4s, v16.4s\n"
+ "and v16.16b, v31.16b, v26.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "sshr v31.4s, v31.4s, #0x1f\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sqadd v30.4s, v30.4s, v16.4s\n"
- "sqadd v25.4s, v25.4s, v31.4s\n"
- "sqadd v20.4s, v20.4s, v0.4s\n"
- "srshl v5.4s, v5.4s, v4.4s\n"
- "srshl v30.4s, v30.4s, v4.4s\n"
- "srshl v25.4s, v25.4s, v4.4s\n"
- "srshl v20.4s, v20.4s, v4.4s\n"
- "add v5.4s, v5.4s, v10.4s\n"
- "add v30.4s, v30.4s, v10.4s\n"
- "add v25.4s, v25.4s, v10.4s\n"
- "add v20.4s, v20.4s, v10.4s\n"
- "smax v5.4s, v5.4s, v13.4s\n"
+ "sqrdmulh v30.4s, v30.4s, v20.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v20.4s\n"
+ "sqrdmulh v28.4s, v28.4s, v20.4s\n"
+ "sqadd v31.4s, v31.4s, v16.4s\n"
+ "and v18.16b, v30.16b, v26.16b\n"
+ "and v17.16b, v29.16b, v26.16b\n"
+ "and v16.16b, v28.16b, v26.16b\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v18.4s\n"
+ "sqadd v29.4s, v29.4s, v17.4s\n"
+ "sqadd v28.4s, v28.4s, v16.4s\n"
+ "srshl v31.4s, v31.4s, v26.4s\n"
+ "srshl v30.4s, v30.4s, v26.4s\n"
+ "srshl v29.4s, v29.4s, v26.4s\n"
+ "srshl v28.4s, v28.4s, v26.4s\n"
+ "add v31.4s, v31.4s, v14.4s\n"
+ "add v30.4s, v30.4s, v14.4s\n"
+ "add v29.4s, v29.4s, v14.4s\n"
+ "add v28.4s, v28.4s, v14.4s\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
"smax v30.4s, v30.4s, v13.4s\n"
- "smax v25.4s, v25.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v13.4s\n"
- "smin v5.4s, v5.4s, v12.4s\n"
- "smin v30.4s, v30.4s, v12.4s\n"
- "smin v25.4s, v25.4s, v12.4s\n"
- "smin v20.4s, v20.4s, v12.4s\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smin v31.4s, v31.4s, v11.4s\n"
+ "smin v30.4s, v30.4s, v11.4s\n"
+ "smin v29.4s, v29.4s, v11.4s\n"
+ "smin v28.4s, v28.4s, v11.4s\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
- "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "uzp1 v31.16b, v31.16b, v31.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v29.16b, v29.16b, v29.16b\n"
+ "uzp1 v28.16b, v28.16b, v28.16b\n"
"32:" // Oddments: Unroll 3: Oddment store
+ "add x25, x25, x27\n"
"add x24, x24, x27\n"
"add x23, x23, x27\n"
"add x22, x22, x27\n"
- "add x21, x21, x27\n"
"tbz x20, #1, 33f\n"
- "st1 { v5.h }[0], [x24], #0x2\n"
- "st1 { v30.h }[0], [x23], #0x2\n"
- "st1 { v25.h }[0], [x22], #0x2\n"
- "st1 { v20.h }[0], [x21], #0x2\n"
+ "st1 { v31.h }[0], [x25], #0x2\n"
+ "st1 { v30.h }[0], [x24], #0x2\n"
+ "st1 { v29.h }[0], [x23], #0x2\n"
+ "st1 { v28.h }[0], [x22], #0x2\n"
"tbz x20, #0, 34f\n"
- "st1 { v5.b }[2], [x24], #0x1\n"
- "st1 { v30.b }[2], [x23], #0x1\n"
- "st1 { v25.b }[2], [x22], #0x1\n"
- "st1 { v20.b }[2], [x21], #0x1\n"
+ "st1 { v31.b }[2], [x25], #0x1\n"
+ "st1 { v30.b }[2], [x24], #0x1\n"
+ "st1 { v29.b }[2], [x23], #0x1\n"
+ "st1 { v28.b }[2], [x22], #0x1\n"
"b 34f\n"
"33:" // Oddments: Unroll 3: Oddment store: Bit 1: Unset
- "st1 { v5.b }[0], [x24], #0x1\n"
- "st1 { v30.b }[0], [x23], #0x1\n"
- "st1 { v25.b }[0], [x22], #0x1\n"
- "st1 { v20.b }[0], [x21], #0x1\n"
+ "st1 { v31.b }[0], [x25], #0x1\n"
+ "st1 { v30.b }[0], [x24], #0x1\n"
+ "st1 { v29.b }[0], [x23], #0x1\n"
+ "st1 { v28.b }[0], [x22], #0x1\n"
"34:" // Oddments: Unroll 3: Oddment store: Bit 1: End
"35:" // End
: [params] "+&r" (params)
: [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [outptrs] "r" (outptrs), [qp] "r" (&qp)
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}