aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp661
1 files changed, 330 insertions, 331 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp
index 6ca3976f02..3426fbc3f9 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp
@@ -120,9 +120,9 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
"add x13, x15, x2\n"
"add x5, x5, x20, LSL #2\n" // outptrs[0] += offset * sizeof(float)
"add x20, %x[params_struct], %[offsetof_args_min]\n"
- "ld1r { v18.4s }, [x20]\n"
+ "ld1r { v27.4s }, [x20]\n"
"add x20, %x[params_struct], %[offsetof_args_max]\n"
- "ld1r { v17.4s }, [x20]\n"
+ "ld1r { v15.4s }, [x20]\n"
"add x12, x14, x24, LSL #2\n"
"add x11, x13, x2\n"
"add x10, x5, x21, LSL #2\n"
@@ -130,7 +130,7 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
"mov x21, #0x0\n"
"sub x20, XZR, x23\n"
"cbz x22, 4f\n"
- "ldr q16, [x8, #0x0]\n"
+ "ldr q25, [x8, #0x0]\n"
"ldr q0, [x8, #0x10]\n"
"cmp x23, x22, LSL #4\n"
"ldr q1, [x8, #0x20]\n"
@@ -150,366 +150,366 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
"ld1 { v14.4s }, [x17]\n"
"bge 3f\n"
"2:" // Tile loop: Channel loop
- "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n"
- "ldr q5, [x7, x15]\n"
- "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n"
+ "mov v30.16b, v25.16b\n fmla v30.4s, v0.4s, v5.4s\n"
+ "ldr q23, [x7, x15]\n"
+ "mov v31.16b, v25.16b\n fmla v31.4s, v0.4s, v6.4s\n"
"add x23, x23, #0x10\n"
- "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n"
- "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n"
- "ldr q0, [x8, #0x0]\n"
- "ldr q16, [x8, #0x140]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "ldr q6, [x7, x13]\n"
- "fmla v29.4s, v1.4s, v9.4s\n"
+ "mov v29.16b, v25.16b\n fmla v29.4s, v0.4s, v7.4s\n"
+ "mov v28.16b, v25.16b\n fmla v28.4s, v0.4s, v8.4s\n"
+ "ldr q19, [x8, #0x0]\n"
+ "ldr q25, [x8, #0x140]\n"
+ "fmla v30.4s, v1.4s, v6.4s\n"
+ "ldr q21, [x7, x13]\n"
+ "fmla v31.4s, v1.4s, v9.4s\n"
"add x7, x7, #0x10\n"
- "fmla v30.4s, v1.4s, v8.4s\n"
- "fmla v31.4s, v1.4s, v13.4s\n"
+ "fmla v29.4s, v1.4s, v8.4s\n"
+ "fmla v28.4s, v1.4s, v13.4s\n"
"ldr q1, [x8, #0x10]\n"
"cmp x23, x22, LSL #4\n"
- "fmla v28.4s, v2.4s, v9.4s\n"
- "ldr q9, [x4, x11]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
+ "fmla v30.4s, v2.4s, v9.4s\n"
+ "ldr q18, [x4, x11]\n"
+ "fmla v31.4s, v2.4s, v11.4s\n"
"add x4, x4, #0x10\n"
- "fmla v30.4s, v2.4s, v13.4s\n"
- "fmla v31.4s, v2.4s, v5.4s\n"
- "ldr q2, [x8, #0x20]\n"
+ "fmla v29.4s, v2.4s, v13.4s\n"
+ "fmla v28.4s, v2.4s, v23.4s\n"
+ "ldr q17, [x8, #0x20]\n"
"add x20, x20, #0x10\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x17, x2]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
+ "fmla v30.4s, v3.4s, v11.4s\n"
+ "ldr q6, [x17, x2]\n"
+ "fmla v31.4s, v3.4s, v12.4s\n"
"add x21, x21, #0x10\n"
- "fmla v30.4s, v3.4s, v5.4s\n"
- "fmla v31.4s, v3.4s, v6.4s\n"
- "ldr q3, [x8, #0x30]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x17, x6]\n"
- "fmla v29.4s, v4.4s, v9.4s\n"
- "ldr q9, [x17, x15]\n"
- "fmla v30.4s, v4.4s, v6.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x8, #0x40]\n"
- "fmla v28.4s, v0.4s, v7.4s\n"
+ "fmla v29.4s, v3.4s, v23.4s\n"
+ "fmla v28.4s, v3.4s, v21.4s\n"
+ "ldr q16, [x8, #0x30]\n"
+ "fmla v30.4s, v4.4s, v12.4s\n"
+ "ldr q2, [x17, x6]\n"
+ "fmla v31.4s, v4.4s, v18.4s\n"
+ "ldr q0, [x17, x15]\n"
+ "fmla v29.4s, v4.4s, v21.4s\n"
+ "fmla v28.4s, v4.4s, v10.4s\n"
+ "ldr q20, [x8, #0x40]\n"
+ "fmla v30.4s, v19.4s, v7.4s\n"
"ld1 { v7.4s }, [x7]\n"
- "fmla v29.4s, v0.4s, v8.4s\n"
- "fmla v30.4s, v0.4s, v14.4s\n"
- "fmla v31.4s, v0.4s, v11.4s\n"
- "ldr q0, [x8, #0x50]\n"
- "fmla v28.4s, v1.4s, v8.4s\n"
- "ldr q8, [x17, x11]\n"
- "fmla v29.4s, v1.4s, v13.4s\n"
- "fmla v30.4s, v1.4s, v11.4s\n"
- "fmla v31.4s, v1.4s, v12.4s\n"
- "ldr q1, [x8, #0x60]\n"
- "fmla v28.4s, v2.4s, v13.4s\n"
- "ldr q13, [x17, x13]\n"
- "fmla v29.4s, v2.4s, v5.4s\n"
+ "fmla v31.4s, v19.4s, v8.4s\n"
+ "fmla v29.4s, v19.4s, v14.4s\n"
+ "fmla v28.4s, v19.4s, v6.4s\n"
+ "ldr q19, [x8, #0x50]\n"
+ "fmla v30.4s, v1.4s, v8.4s\n"
+ "ldr q26, [x17, x11]\n"
+ "fmla v31.4s, v1.4s, v13.4s\n"
+ "fmla v29.4s, v1.4s, v6.4s\n"
+ "fmla v28.4s, v1.4s, v2.4s\n"
+ "ldr q18, [x8, #0x60]\n"
+ "fmla v30.4s, v17.4s, v13.4s\n"
+ "ldr q1, [x17, x13]\n"
+ "fmla v31.4s, v17.4s, v23.4s\n"
"add x17, x17, #0x10\n"
- "fmla v30.4s, v2.4s, v12.4s\n"
- "fmla v31.4s, v2.4s, v9.4s\n"
- "ldr q2, [x8, #0x70]\n"
- "fmla v28.4s, v3.4s, v5.4s\n"
- "ld1 { v5.4s }, [x16]\n"
- "fmla v29.4s, v3.4s, v6.4s\n"
- "fmla v30.4s, v3.4s, v9.4s\n"
- "fmla v31.4s, v3.4s, v13.4s\n"
- "ldr q3, [x8, #0x80]\n"
- "fmla v28.4s, v4.4s, v6.4s\n"
- "ldr q6, [x16, x2]\n"
- "fmla v29.4s, v4.4s, v10.4s\n"
- "ldr q10, [x16, x6]\n"
- "fmla v30.4s, v4.4s, v13.4s\n"
- "fmla v31.4s, v4.4s, v8.4s\n"
- "ldr q4, [x8, #0x90]\n"
- "fmla v28.4s, v0.4s, v14.4s\n"
- "ldr q14, [x16, x11]\n"
- "fmla v29.4s, v0.4s, v11.4s\n"
- "fmla v30.4s, v0.4s, v5.4s\n"
- "fmla v31.4s, v0.4s, v6.4s\n"
- "ldr q0, [x8, #0xa0]\n"
- "fmla v28.4s, v1.4s, v11.4s\n"
- "ldr q11, [x16, x15]\n"
- "fmla v29.4s, v1.4s, v12.4s\n"
- "fmla v30.4s, v1.4s, v6.4s\n"
- "fmla v31.4s, v1.4s, v10.4s\n"
- "ldr q1, [x8, #0xb0]\n"
- "fmla v28.4s, v2.4s, v12.4s\n"
- "ldr q12, [x16, x13]\n"
- "fmla v29.4s, v2.4s, v9.4s\n"
+ "fmla v29.4s, v17.4s, v2.4s\n"
+ "fmla v28.4s, v17.4s, v0.4s\n"
+ "ldr q17, [x8, #0x70]\n"
+ "fmla v30.4s, v16.4s, v23.4s\n"
+ "ld1 { v24.4s }, [x16]\n"
+ "fmla v31.4s, v16.4s, v21.4s\n"
+ "fmla v29.4s, v16.4s, v0.4s\n"
+ "fmla v28.4s, v16.4s, v1.4s\n"
+ "ldr q16, [x8, #0x80]\n"
+ "fmla v30.4s, v20.4s, v21.4s\n"
+ "ldr q23, [x16, x2]\n"
+ "fmla v31.4s, v20.4s, v10.4s\n"
+ "ldr q22, [x16, x6]\n"
+ "fmla v29.4s, v20.4s, v1.4s\n"
+ "fmla v28.4s, v20.4s, v26.4s\n"
+ "ldr q21, [x8, #0x90]\n"
+ "fmla v30.4s, v19.4s, v14.4s\n"
+ "ldr q5, [x16, x11]\n"
+ "fmla v31.4s, v19.4s, v6.4s\n"
+ "fmla v29.4s, v19.4s, v24.4s\n"
+ "fmla v28.4s, v19.4s, v23.4s\n"
+ "ldr q11, [x8, #0xa0]\n"
+ "fmla v30.4s, v18.4s, v6.4s\n"
+ "ldr q20, [x16, x15]\n"
+ "fmla v31.4s, v18.4s, v2.4s\n"
+ "fmla v29.4s, v18.4s, v23.4s\n"
+ "fmla v28.4s, v18.4s, v22.4s\n"
+ "ldr q18, [x8, #0xb0]\n"
+ "fmla v30.4s, v17.4s, v2.4s\n"
+ "ldr q19, [x16, x13]\n"
+ "fmla v31.4s, v17.4s, v0.4s\n"
"add x16, x16, #0x10\n"
- "fmla v30.4s, v2.4s, v10.4s\n"
- "fmla v31.4s, v2.4s, v11.4s\n"
- "ldr q2, [x8, #0xc0]\n"
- "fmla v28.4s, v3.4s, v9.4s\n"
- "ld1 { v9.4s }, [x14]\n"
- "fmla v29.4s, v3.4s, v13.4s\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v12.4s\n"
- "ldr q3, [x8, #0xd0]\n"
- "fmla v28.4s, v4.4s, v13.4s\n"
- "ldr q13, [x14, x2]\n"
- "fmla v29.4s, v4.4s, v8.4s\n"
- "ldr q8, [x14, x13]\n"
- "fmla v30.4s, v4.4s, v12.4s\n"
- "fmla v31.4s, v4.4s, v14.4s\n"
- "ldr q4, [x8, #0xe0]\n"
- "fmla v28.4s, v0.4s, v5.4s\n"
- "ldr q5, [x14, x6]\n"
- "fmla v29.4s, v0.4s, v6.4s\n"
- "fmla v30.4s, v0.4s, v9.4s\n"
- "fmla v31.4s, v0.4s, v13.4s\n"
- "ldr q0, [x8, #0xf0]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "ldr q6, [x14, x15]\n"
- "fmla v29.4s, v1.4s, v10.4s\n"
- "fmla v30.4s, v1.4s, v13.4s\n"
- "fmla v31.4s, v1.4s, v5.4s\n"
- "ldr q1, [x8, #0x100]\n"
- "fmla v28.4s, v2.4s, v10.4s\n"
- "ldr q10, [x14, x11]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
+ "fmla v29.4s, v17.4s, v22.4s\n"
+ "fmla v28.4s, v17.4s, v20.4s\n"
+ "ldr q17, [x8, #0xc0]\n"
+ "fmla v30.4s, v16.4s, v0.4s\n"
+ "ld1 { v0.4s }, [x14]\n"
+ "fmla v31.4s, v16.4s, v1.4s\n"
+ "fmla v29.4s, v16.4s, v20.4s\n"
+ "fmla v28.4s, v16.4s, v19.4s\n"
+ "ldr q16, [x8, #0xd0]\n"
+ "fmla v30.4s, v21.4s, v1.4s\n"
+ "ldr q4, [x14, x2]\n"
+ "fmla v31.4s, v21.4s, v26.4s\n"
+ "ldr q12, [x14, x13]\n"
+ "fmla v29.4s, v21.4s, v19.4s\n"
+ "fmla v28.4s, v21.4s, v5.4s\n"
+ "ldr q13, [x8, #0xe0]\n"
+ "fmla v30.4s, v11.4s, v24.4s\n"
+ "ldr q6, [x14, x6]\n"
+ "fmla v31.4s, v11.4s, v23.4s\n"
+ "fmla v29.4s, v11.4s, v0.4s\n"
+ "fmla v28.4s, v11.4s, v4.4s\n"
+ "ldr q24, [x8, #0xf0]\n"
+ "fmla v30.4s, v18.4s, v23.4s\n"
+ "ldr q26, [x14, x15]\n"
+ "fmla v31.4s, v18.4s, v22.4s\n"
+ "fmla v29.4s, v18.4s, v4.4s\n"
+ "fmla v28.4s, v18.4s, v6.4s\n"
+ "ldr q23, [x8, #0x100]\n"
+ "fmla v30.4s, v17.4s, v22.4s\n"
+ "ldr q22, [x14, x11]\n"
+ "fmla v31.4s, v17.4s, v20.4s\n"
"add x14, x14, #0x10\n"
- "fmla v30.4s, v2.4s, v5.4s\n"
- "fmla v31.4s, v2.4s, v6.4s\n"
- "ldr q2, [x8, #0x110]\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ld1 { v11.4s }, [x12]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "fmla v30.4s, v3.4s, v6.4s\n"
- "fmla v31.4s, v3.4s, v8.4s\n"
- "ldr q3, [x8, #0x120]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x12, x2]\n"
- "fmla v29.4s, v4.4s, v14.4s\n"
+ "fmla v29.4s, v17.4s, v6.4s\n"
+ "fmla v28.4s, v17.4s, v26.4s\n"
+ "ldr q21, [x8, #0x110]\n"
+ "fmla v30.4s, v16.4s, v20.4s\n"
+ "ld1 { v18.4s }, [x12]\n"
+ "fmla v31.4s, v16.4s, v19.4s\n"
+ "fmla v29.4s, v16.4s, v26.4s\n"
+ "fmla v28.4s, v16.4s, v12.4s\n"
+ "ldr q20, [x8, #0x120]\n"
+ "fmla v30.4s, v13.4s, v19.4s\n"
+ "ldr q17, [x12, x2]\n"
+ "fmla v31.4s, v13.4s, v5.4s\n"
"ld1 { v14.4s }, [x17]\n"
- "fmla v30.4s, v4.4s, v8.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x8, #0x130]\n"
- "fmla v28.4s, v0.4s, v9.4s\n"
- "ldr q9, [x12, x6]\n"
- "fmla v29.4s, v0.4s, v13.4s\n"
- "fmla v30.4s, v0.4s, v11.4s\n"
- "ldr q11, [x12, x15]\n"
- "fmla v31.4s, v0.4s, v12.4s\n"
+ "fmla v29.4s, v13.4s, v12.4s\n"
+ "fmla v28.4s, v13.4s, v22.4s\n"
+ "ldr q19, [x8, #0x130]\n"
+ "fmla v30.4s, v24.4s, v0.4s\n"
+ "ldr q16, [x12, x6]\n"
+ "fmla v31.4s, v24.4s, v4.4s\n"
+ "fmla v29.4s, v24.4s, v18.4s\n"
+ "ldr q18, [x12, x15]\n"
+ "fmla v28.4s, v24.4s, v17.4s\n"
"ldr q0, [x8, #0x150]\n"
- "fmla v28.4s, v1.4s, v13.4s\n"
+ "fmla v30.4s, v23.4s, v4.4s\n"
"ldr q13, [x7, x6]\n"
- "fmla v29.4s, v1.4s, v5.4s\n"
- "fmla v30.4s, v1.4s, v12.4s\n"
- "ldr q12, [x12, x13]\n"
- "fmla v31.4s, v1.4s, v9.4s\n"
+ "fmla v31.4s, v23.4s, v6.4s\n"
+ "fmla v29.4s, v23.4s, v17.4s\n"
+ "ldr q17, [x12, x13]\n"
+ "fmla v28.4s, v23.4s, v16.4s\n"
"ldr q1, [x8, #0x160]\n"
- "fmla v28.4s, v2.4s, v5.4s\n"
+ "fmla v30.4s, v21.4s, v6.4s\n"
"ld1 { v5.4s }, [x4]\n"
- "fmla v29.4s, v2.4s, v6.4s\n"
- "fmla v30.4s, v2.4s, v9.4s\n"
- "ldr q9, [x12, x11]\n"
- "fmla v31.4s, v2.4s, v11.4s\n"
+ "fmla v31.4s, v21.4s, v26.4s\n"
+ "fmla v29.4s, v21.4s, v16.4s\n"
+ "ldr q16, [x12, x11]\n"
+ "fmla v28.4s, v21.4s, v18.4s\n"
"ldr q2, [x8, #0x170]\n"
- "fmla v28.4s, v3.4s, v6.4s\n"
+ "fmla v30.4s, v20.4s, v26.4s\n"
"ldr q6, [x4, x2]\n"
- "fmla v29.4s, v3.4s, v8.4s\n"
+ "fmla v31.4s, v20.4s, v12.4s\n"
"add x12, x12, #0x10\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
+ "fmla v29.4s, v20.4s, v18.4s\n"
"ldr q11, [x4, x15]\n"
- "fmla v31.4s, v3.4s, v12.4s\n"
+ "fmla v28.4s, v20.4s, v17.4s\n"
"ldr q3, [x8, #0x180]\n"
- "fmla v28.4s, v4.4s, v8.4s\n"
+ "fmla v30.4s, v19.4s, v12.4s\n"
"ldr q8, [x7, x2]\n"
- "fmla v29.4s, v4.4s, v10.4s\n"
+ "fmla v31.4s, v19.4s, v22.4s\n"
"ldr q10, [x7, x11]\n"
- "fmla v30.4s, v4.4s, v12.4s\n"
+ "fmla v29.4s, v19.4s, v17.4s\n"
"ldr q12, [x4, x13]\n"
- "fmla v31.4s, v4.4s, v9.4s\n"
+ "fmla v28.4s, v19.4s, v16.4s\n"
"ldr q9, [x4, x6]\n"
"ldr q4, [x8, #0x190]\n"
- "fmax v28.4s, v28.4s, v18.4s\n"
- "fmax v29.4s, v29.4s, v18.4s\n"
+ "fmax v30.4s, v30.4s, v27.4s\n"
+ "fmax v31.4s, v31.4s, v27.4s\n"
"add x8, x8, #0x1a0\n"
- "fmax v30.4s, v30.4s, v18.4s\n"
- "fmax v31.4s, v31.4s, v18.4s\n"
- "fmin v28.4s, v28.4s, v17.4s\n"
- "fmin v29.4s, v29.4s, v17.4s\n"
- "st1 { v28.4s }, [x5]\n"
- "fmin v30.4s, v30.4s, v17.4s\n"
- "fmin v31.4s, v31.4s, v17.4s\n"
- "str q29, [x5, x3]\n"
+ "fmax v29.4s, v29.4s, v27.4s\n"
+ "fmax v28.4s, v28.4s, v27.4s\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
+ "st1 { v30.4s }, [x5]\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
+ "fmin v28.4s, v28.4s, v15.4s\n"
+ "str q31, [x5, x3]\n"
"add x5, x5, #0x10\n"
- "st1 { v30.4s }, [x10]\n"
- "str q31, [x10, x3]\n"
+ "st1 { v29.4s }, [x10]\n"
+ "str q28, [x10, x3]\n"
"add x10, x10, #0x10\n"
"blt 2b\n"
"3:" // Tile loop: Channel tail
- "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n"
- "ldr q5, [x7, x15]\n"
- "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n"
- "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n"
- "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n"
- "ldr q0, [x8, #0x0]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "ldr q6, [x7, x13]\n"
- "fmla v29.4s, v1.4s, v9.4s\n"
+ "mov v31.16b, v25.16b\n fmla v31.4s, v0.4s, v5.4s\n"
+ "ldr q22, [x7, x15]\n"
+ "mov v5.16b, v25.16b\n fmla v5.4s, v0.4s, v6.4s\n"
+ "mov v30.16b, v25.16b\n fmla v30.4s, v0.4s, v7.4s\n"
+ "mov v29.16b, v25.16b\n fmla v29.4s, v0.4s, v8.4s\n"
+ "ldr q19, [x8, #0x0]\n"
+ "fmla v31.4s, v1.4s, v6.4s\n"
+ "ldr q21, [x7, x13]\n"
+ "fmla v5.4s, v1.4s, v9.4s\n"
"add x7, x7, #0x10\n"
"fmla v30.4s, v1.4s, v8.4s\n"
- "fmla v31.4s, v1.4s, v13.4s\n"
- "ldr q1, [x8, #0x10]\n"
- "fmla v28.4s, v2.4s, v9.4s\n"
- "ldr q9, [x4, x11]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
- "add x4, x4, #0x10\n"
- "fmla v30.4s, v2.4s, v13.4s\n"
- "fmla v31.4s, v2.4s, v5.4s\n"
- "ldr q2, [x8, #0x20]\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x17, x2]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "fmla v30.4s, v3.4s, v5.4s\n"
- "fmla v31.4s, v3.4s, v6.4s\n"
- "ldr q3, [x8, #0x30]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x17, x6]\n"
- "fmla v29.4s, v4.4s, v9.4s\n"
- "ldr q9, [x17, x15]\n"
- "fmla v30.4s, v4.4s, v6.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x8, #0x40]\n"
- "fmla v28.4s, v0.4s, v7.4s\n"
- "fmla v29.4s, v0.4s, v8.4s\n"
- "fmla v30.4s, v0.4s, v14.4s\n"
- "fmla v31.4s, v0.4s, v11.4s\n"
- "ldr q0, [x8, #0x50]\n"
- "fmla v28.4s, v1.4s, v8.4s\n"
- "ldr q8, [x17, x11]\n"
"fmla v29.4s, v1.4s, v13.4s\n"
- "fmla v30.4s, v1.4s, v11.4s\n"
- "fmla v31.4s, v1.4s, v12.4s\n"
- "ldr q1, [x8, #0x60]\n"
- "fmla v28.4s, v2.4s, v13.4s\n"
- "ldr q13, [x17, x13]\n"
- "fmla v29.4s, v2.4s, v5.4s\n"
- "add x17, x17, #0x10\n"
- "fmla v30.4s, v2.4s, v12.4s\n"
+ "ldr q18, [x8, #0x10]\n"
"fmla v31.4s, v2.4s, v9.4s\n"
- "ldr q2, [x8, #0x70]\n"
- "fmla v28.4s, v3.4s, v5.4s\n"
- "ld1 { v5.4s }, [x16]\n"
- "fmla v29.4s, v3.4s, v6.4s\n"
- "fmla v30.4s, v3.4s, v9.4s\n"
- "fmla v31.4s, v3.4s, v13.4s\n"
- "ldr q3, [x8, #0x80]\n"
- "fmla v28.4s, v4.4s, v6.4s\n"
- "ldr q6, [x16, x2]\n"
+ "ldr q16, [x4, x11]\n"
+ "fmla v5.4s, v2.4s, v11.4s\n"
+ "add x4, x4, #0x10\n"
+ "fmla v30.4s, v2.4s, v13.4s\n"
+ "fmla v29.4s, v2.4s, v22.4s\n"
+ "ldr q17, [x8, #0x20]\n"
+ "fmla v31.4s, v3.4s, v11.4s\n"
+ "ldr q6, [x17, x2]\n"
+ "fmla v5.4s, v3.4s, v12.4s\n"
+ "fmla v30.4s, v3.4s, v22.4s\n"
+ "fmla v29.4s, v3.4s, v21.4s\n"
+ "ldr q20, [x8, #0x30]\n"
+ "fmla v31.4s, v4.4s, v12.4s\n"
+ "ldr q2, [x17, x6]\n"
+ "fmla v5.4s, v4.4s, v16.4s\n"
+ "ldr q28, [x17, x15]\n"
+ "fmla v30.4s, v4.4s, v21.4s\n"
"fmla v29.4s, v4.4s, v10.4s\n"
- "ldr q10, [x16, x6]\n"
- "fmla v30.4s, v4.4s, v13.4s\n"
- "fmla v31.4s, v4.4s, v8.4s\n"
- "ldr q4, [x8, #0x90]\n"
- "fmla v28.4s, v0.4s, v14.4s\n"
- "ldr q14, [x16, x11]\n"
- "fmla v29.4s, v0.4s, v11.4s\n"
- "fmla v30.4s, v0.4s, v5.4s\n"
- "fmla v31.4s, v0.4s, v6.4s\n"
- "ldr q0, [x8, #0xa0]\n"
- "fmla v28.4s, v1.4s, v11.4s\n"
- "ldr q11, [x16, x15]\n"
- "fmla v29.4s, v1.4s, v12.4s\n"
- "fmla v30.4s, v1.4s, v6.4s\n"
- "fmla v31.4s, v1.4s, v10.4s\n"
- "ldr q1, [x8, #0xb0]\n"
- "fmla v28.4s, v2.4s, v12.4s\n"
- "ldr q12, [x16, x13]\n"
- "fmla v29.4s, v2.4s, v9.4s\n"
+ "ldr q16, [x8, #0x40]\n"
+ "fmla v31.4s, v19.4s, v7.4s\n"
+ "fmla v5.4s, v19.4s, v8.4s\n"
+ "fmla v30.4s, v19.4s, v14.4s\n"
+ "fmla v29.4s, v19.4s, v6.4s\n"
+ "ldr q19, [x8, #0x50]\n"
+ "fmla v31.4s, v18.4s, v8.4s\n"
+ "ldr q1, [x17, x11]\n"
+ "fmla v5.4s, v18.4s, v13.4s\n"
+ "fmla v30.4s, v18.4s, v6.4s\n"
+ "fmla v29.4s, v18.4s, v2.4s\n"
+ "ldr q18, [x8, #0x60]\n"
+ "fmla v31.4s, v17.4s, v13.4s\n"
+ "ldr q26, [x17, x13]\n"
+ "fmla v5.4s, v17.4s, v22.4s\n"
+ "add x17, x17, #0x10\n"
+ "fmla v30.4s, v17.4s, v2.4s\n"
+ "fmla v29.4s, v17.4s, v28.4s\n"
+ "ldr q17, [x8, #0x70]\n"
+ "fmla v31.4s, v20.4s, v22.4s\n"
+ "ld1 { v25.4s }, [x16]\n"
+ "fmla v5.4s, v20.4s, v21.4s\n"
+ "fmla v30.4s, v20.4s, v28.4s\n"
+ "fmla v29.4s, v20.4s, v26.4s\n"
+ "ldr q24, [x8, #0x80]\n"
+ "fmla v31.4s, v16.4s, v21.4s\n"
+ "ldr q23, [x16, x2]\n"
+ "fmla v5.4s, v16.4s, v10.4s\n"
+ "ldr q0, [x16, x6]\n"
+ "fmla v30.4s, v16.4s, v26.4s\n"
+ "fmla v29.4s, v16.4s, v1.4s\n"
+ "ldr q22, [x8, #0x90]\n"
+ "fmla v31.4s, v19.4s, v14.4s\n"
+ "ldr q16, [x16, x11]\n"
+ "fmla v5.4s, v19.4s, v6.4s\n"
+ "fmla v30.4s, v19.4s, v25.4s\n"
+ "fmla v29.4s, v19.4s, v23.4s\n"
+ "ldr q21, [x8, #0xa0]\n"
+ "fmla v31.4s, v18.4s, v6.4s\n"
+ "ldr q20, [x16, x15]\n"
+ "fmla v5.4s, v18.4s, v2.4s\n"
+ "fmla v30.4s, v18.4s, v23.4s\n"
+ "fmla v29.4s, v18.4s, v0.4s\n"
+ "ldr q18, [x8, #0xb0]\n"
+ "fmla v31.4s, v17.4s, v2.4s\n"
+ "ldr q19, [x16, x13]\n"
+ "fmla v5.4s, v17.4s, v28.4s\n"
"add x16, x16, #0x10\n"
- "fmla v30.4s, v2.4s, v10.4s\n"
- "fmla v31.4s, v2.4s, v11.4s\n"
- "ldr q2, [x8, #0xc0]\n"
- "fmla v28.4s, v3.4s, v9.4s\n"
- "ld1 { v9.4s }, [x14]\n"
- "fmla v29.4s, v3.4s, v13.4s\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v12.4s\n"
- "ldr q3, [x8, #0xd0]\n"
- "fmla v28.4s, v4.4s, v13.4s\n"
- "ldr q13, [x14, x2]\n"
- "fmla v29.4s, v4.4s, v8.4s\n"
- "ldr q8, [x14, x13]\n"
- "fmla v30.4s, v4.4s, v12.4s\n"
- "fmla v31.4s, v4.4s, v14.4s\n"
- "ldr q4, [x8, #0xe0]\n"
- "fmla v28.4s, v0.4s, v5.4s\n"
- "ldr q5, [x14, x6]\n"
- "fmla v29.4s, v0.4s, v6.4s\n"
- "fmla v30.4s, v0.4s, v9.4s\n"
- "fmla v31.4s, v0.4s, v13.4s\n"
- "ldr q0, [x8, #0xf0]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "ldr q6, [x14, x15]\n"
- "fmla v29.4s, v1.4s, v10.4s\n"
- "fmla v30.4s, v1.4s, v13.4s\n"
- "fmla v31.4s, v1.4s, v5.4s\n"
- "ldr q1, [x8, #0x100]\n"
- "fmla v28.4s, v2.4s, v10.4s\n"
- "ldr q10, [x14, x11]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
+ "fmla v30.4s, v17.4s, v0.4s\n"
+ "fmla v29.4s, v17.4s, v20.4s\n"
+ "ldr q17, [x8, #0xc0]\n"
+ "fmla v31.4s, v24.4s, v28.4s\n"
+ "ld1 { v7.4s }, [x14]\n"
+ "fmla v5.4s, v24.4s, v26.4s\n"
+ "fmla v30.4s, v24.4s, v20.4s\n"
+ "fmla v29.4s, v24.4s, v19.4s\n"
+ "ldr q2, [x8, #0xd0]\n"
+ "fmla v31.4s, v22.4s, v26.4s\n"
+ "ldr q28, [x14, x2]\n"
+ "fmla v5.4s, v22.4s, v1.4s\n"
+ "ldr q13, [x14, x13]\n"
+ "fmla v30.4s, v22.4s, v19.4s\n"
+ "fmla v29.4s, v22.4s, v16.4s\n"
+ "ldr q14, [x8, #0xe0]\n"
+ "fmla v31.4s, v21.4s, v25.4s\n"
+ "ldr q26, [x14, x6]\n"
+ "fmla v5.4s, v21.4s, v23.4s\n"
+ "fmla v30.4s, v21.4s, v7.4s\n"
+ "fmla v29.4s, v21.4s, v28.4s\n"
+ "ldr q25, [x8, #0xf0]\n"
+ "fmla v31.4s, v18.4s, v23.4s\n"
+ "ldr q24, [x14, x15]\n"
+ "fmla v5.4s, v18.4s, v0.4s\n"
+ "fmla v30.4s, v18.4s, v28.4s\n"
+ "fmla v29.4s, v18.4s, v26.4s\n"
+ "ldr q23, [x8, #0x100]\n"
+ "fmla v31.4s, v17.4s, v0.4s\n"
+ "ldr q22, [x14, x11]\n"
+ "fmla v5.4s, v17.4s, v20.4s\n"
"add x14, x14, #0x10\n"
- "fmla v30.4s, v2.4s, v5.4s\n"
- "fmla v31.4s, v2.4s, v6.4s\n"
- "ldr q2, [x8, #0x110]\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ld1 { v11.4s }, [x12]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "fmla v30.4s, v3.4s, v6.4s\n"
- "fmla v31.4s, v3.4s, v8.4s\n"
- "ldr q3, [x8, #0x120]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x12, x2]\n"
- "fmla v29.4s, v4.4s, v14.4s\n"
- "fmla v30.4s, v4.4s, v8.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x8, #0x130]\n"
+ "fmla v30.4s, v17.4s, v26.4s\n"
+ "fmla v29.4s, v17.4s, v24.4s\n"
+ "ldr q21, [x8, #0x110]\n"
+ "fmla v31.4s, v2.4s, v20.4s\n"
+ "ld1 { v18.4s }, [x12]\n"
+ "fmla v5.4s, v2.4s, v19.4s\n"
+ "fmla v30.4s, v2.4s, v24.4s\n"
+ "fmla v29.4s, v2.4s, v13.4s\n"
+ "ldr q20, [x8, #0x120]\n"
+ "fmla v31.4s, v14.4s, v19.4s\n"
+ "ldr q17, [x12, x2]\n"
+ "fmla v5.4s, v14.4s, v16.4s\n"
+ "fmla v30.4s, v14.4s, v13.4s\n"
+ "fmla v29.4s, v14.4s, v22.4s\n"
+ "ldr q19, [x8, #0x130]\n"
"add x8, x8, #0x140\n"
- "fmla v28.4s, v0.4s, v9.4s\n"
- "ldr q9, [x12, x6]\n"
- "fmla v29.4s, v0.4s, v13.4s\n"
- "fmla v30.4s, v0.4s, v11.4s\n"
- "ldr q11, [x12, x15]\n"
- "fmla v31.4s, v0.4s, v12.4s\n"
- "fmla v28.4s, v1.4s, v13.4s\n"
- "fmla v29.4s, v1.4s, v5.4s\n"
- "fmla v30.4s, v1.4s, v12.4s\n"
- "ldr q12, [x12, x13]\n"
- "fmla v31.4s, v1.4s, v9.4s\n"
- "fmla v28.4s, v2.4s, v5.4s\n"
- "fmla v29.4s, v2.4s, v6.4s\n"
- "fmla v30.4s, v2.4s, v9.4s\n"
- "ldr q9, [x12, x11]\n"
- "fmla v31.4s, v2.4s, v11.4s\n"
+ "fmla v31.4s, v25.4s, v7.4s\n"
+ "ldr q16, [x12, x6]\n"
+ "fmla v5.4s, v25.4s, v28.4s\n"
+ "fmla v30.4s, v25.4s, v18.4s\n"
+ "ldr q18, [x12, x15]\n"
+ "fmla v29.4s, v25.4s, v17.4s\n"
+ "fmla v31.4s, v23.4s, v28.4s\n"
+ "fmla v5.4s, v23.4s, v26.4s\n"
+ "fmla v30.4s, v23.4s, v17.4s\n"
+ "ldr q17, [x12, x13]\n"
+ "fmla v29.4s, v23.4s, v16.4s\n"
+ "fmla v31.4s, v21.4s, v26.4s\n"
+ "fmla v5.4s, v21.4s, v24.4s\n"
+ "fmla v30.4s, v21.4s, v16.4s\n"
+ "ldr q16, [x12, x11]\n"
+ "fmla v29.4s, v21.4s, v18.4s\n"
"add x12, x12, #0x10\n"
- "fmla v28.4s, v3.4s, v6.4s\n"
- "fmla v29.4s, v3.4s, v8.4s\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v12.4s\n"
- "fmla v28.4s, v4.4s, v8.4s\n"
- "fmla v29.4s, v4.4s, v10.4s\n"
- "fmax v28.4s, v28.4s, v18.4s\n"
- "fmla v30.4s, v4.4s, v12.4s\n"
- "fmla v31.4s, v4.4s, v9.4s\n"
- "fmax v29.4s, v29.4s, v18.4s\n"
- "fmax v30.4s, v30.4s, v18.4s\n"
- "fmax v31.4s, v31.4s, v18.4s\n"
- "fmin v28.4s, v28.4s, v17.4s\n"
- "fmin v29.4s, v29.4s, v17.4s\n"
- "st1 { v28.4s }, [x5]\n"
- "fmin v30.4s, v30.4s, v17.4s\n"
- "fmin v31.4s, v31.4s, v17.4s\n"
- "str q29, [x5, x3]\n"
+ "fmla v31.4s, v20.4s, v24.4s\n"
+ "fmla v5.4s, v20.4s, v13.4s\n"
+ "fmla v30.4s, v20.4s, v18.4s\n"
+ "fmla v29.4s, v20.4s, v17.4s\n"
+ "fmla v31.4s, v19.4s, v13.4s\n"
+ "fmla v5.4s, v19.4s, v22.4s\n"
+ "fmax v31.4s, v31.4s, v27.4s\n"
+ "fmla v30.4s, v19.4s, v17.4s\n"
+ "fmla v29.4s, v19.4s, v16.4s\n"
+ "fmax v5.4s, v5.4s, v27.4s\n"
+ "fmax v30.4s, v30.4s, v27.4s\n"
+ "fmax v29.4s, v29.4s, v27.4s\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
+ "fmin v5.4s, v5.4s, v15.4s\n"
+ "st1 { v31.4s }, [x5]\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
+ "str q5, [x5, x3]\n"
"add x5, x5, #0x10\n"
"st1 { v30.4s }, [x10]\n"
- "str q31, [x10, x3]\n"
+ "str q29, [x10, x3]\n"
"add x10, x10, #0x10\n"
"4:" // Tile loop: Oddments
"tst %x[n_channels], #0x3\n"
"beq 61f\n"
- "ldr q16, [x8, #0x0]\n"
+ "ldr q25, [x8, #0x0]\n"
"ldr q0, [x8, #0x10]\n"
"add x9, x4, XZR\n"
"add x28, x4, x2\n"
@@ -561,11 +561,11 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
"ldr s10, [x21, #0x0]\n"
"ldr s14, [x20, #0x0]\n"
"6:" // Tile loop: Oddments: Load inputs: (0, 0), (0, 1), (1, 0), (1, 1), (0, 2), (1, 2), (0, 3), (0, 4), (1, 5), (2, 0): Bit 1: End
- "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n"
- "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n"
+ "mov v28.16b, v25.16b\n fmla v28.4s, v0.4s, v5.4s\n"
+ "mov v29.16b, v25.16b\n fmla v29.4s, v0.4s, v6.4s\n"
"add x20, x7, x15\n"
- "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n"
- "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n"
+ "mov v30.16b, v25.16b\n fmla v30.4s, v0.4s, v7.4s\n"
+ "mov v31.16b, v25.16b\n fmla v31.4s, v0.4s, v8.4s\n"
"fmla v28.4s, v1.4s, v6.4s\n"
"fmla v29.4s, v1.4s, v9.4s\n"
"fmla v30.4s, v1.4s, v8.4s\n"
@@ -934,14 +934,14 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
"ldr s9, [x20, #0x0]\n"
"58:" // Tile loop: Oddments: Load inputs: (5, 5): Bit 1: End
"fmla v31.4s, v4.4s, v9.4s\n"
- "fmax v28.4s, v28.4s, v18.4s\n"
- "fmax v29.4s, v29.4s, v18.4s\n"
- "fmax v30.4s, v30.4s, v18.4s\n"
- "fmax v31.4s, v31.4s, v18.4s\n"
- "fmin v28.4s, v28.4s, v17.4s\n"
- "fmin v29.4s, v29.4s, v17.4s\n"
- "fmin v30.4s, v30.4s, v17.4s\n"
- "fmin v31.4s, v31.4s, v17.4s\n"
+ "fmax v28.4s, v28.4s, v27.4s\n"
+ "fmax v29.4s, v29.4s, v27.4s\n"
+ "fmax v30.4s, v30.4s, v27.4s\n"
+ "fmax v31.4s, v31.4s, v27.4s\n"
+ "fmin v28.4s, v28.4s, v15.4s\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
"tbz %x[n_channels], #1, 59f\n"
"mov x21, x5\n"
"mov x20, x10\n"
@@ -967,7 +967,6 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
"st1 { v29.s }[0], [x21]\n"
"st1 { v31.s }[0], [x20]\n"
"60:" // Tile loop: Oddments: Store: Bit 1: End
-
"61:" // Tile loop: End
"ldr x26, [%x[params_struct], %[offsetof_args_tile_j]]\n"
"ldr x27, [%x[params_struct], %[offsetof_args_tile_i]]\n"
@@ -982,7 +981,7 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_direct_impl(
"blt 1b\n"
:
: [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (&params_struct)
- : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v16", "v17", "v18", "v28", "v29", "v30", "v31", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "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", "x2", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}