aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp567
1 files changed, 283 insertions, 284 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
index 2b1dc3646d..4e64a2bf2b 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp
@@ -120,9 +120,9 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"add x9, x11, x8\n"
"add x28, x15, x22, LSL #1\n"
"add x20, %x[params_struct], %[offsetof_args_min]\n"
- "ld1r { v18.8h }, [x20]\n"
+ "ld1r { v15.8h }, [x20]\n"
"add x20, %x[params_struct], %[offsetof_args_max]\n"
- "ld1r { v17.8h }, [x20]\n"
+ "ld1r { v14.8h }, [x20]\n"
"add x27, x10, x25, LSL #1\n"
"add x26, x9, x8\n"
"add x25, x28, x22, LSL #1\n"
@@ -130,7 +130,7 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"mov x21, #0x0\n"
"sub x20, XZR, x24\n"
"cbz x23, 4f\n"
- "ldr q16, [x14, #0x0]\n"
+ "ldr q31, [x14, #0x0]\n"
"ldr q0, [x14, #0x10]\n"
"cmp x24, x23, LSL #4\n"
"ldr q1, [x14, #0x20]\n"
@@ -149,304 +149,304 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"ldr q13, [x13, x11]\n"
"bge 3f\n"
"2:" // Tile loop: Channel loop
- "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n"
- "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n"
+ "mov v29.16b, v31.16b\n fmla v29.8h, v7.8h, v9.8h\n"
+ "mov v28.16b, v31.16b\n fmla v28.8h, v8.8h, v9.8h\n"
"add x24, x24, #0x10\n"
"cmp x24, x23, LSL #4\n"
- "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n"
- "fmla v24.8h, v4.8h, v13.8h\n"
+ "mov v27.16b, v31.16b\n fmla v27.8h, v6.8h, v9.8h\n"
+ "fmla v29.8h, v4.8h, v13.8h\n"
"add x20, x20, #0x10\n"
"add x21, x21, #0x10\n"
- "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n"
- "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n"
- "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n"
- "fmla v23.8h, v0.8h, v10.8h\n"
- "ldr q10, [x12, x9]\n"
- "fmla v25.8h, v2.8h, v11.8h\n"
- "ldr q11, [x12, x8]\n"
- "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n"
- "fmla v24.8h, v6.8h, v11.8h\n"
- "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n"
- "fmla v23.8h, v5.8h, v13.8h\n"
- "fmla v25.8h, v3.8h, v13.8h\n"
+ "mov v26.16b, v31.16b\n fmla v26.8h, v5.8h, v9.8h\n"
+ "mov v25.16b, v31.16b\n fmla v25.8h, v4.8h, v9.8h\n"
+ "mov v24.16b, v31.16b\n fmla v24.8h, v3.8h, v9.8h\n"
+ "fmla v28.8h, v0.8h, v10.8h\n"
+ "ldr q23, [x12, x9]\n"
+ "fmla v27.8h, v2.8h, v11.8h\n"
+ "ldr q18, [x12, x8]\n"
+ "mov v22.16b, v31.16b\n fmla v22.8h, v2.8h, v9.8h\n"
+ "fmla v29.8h, v6.8h, v18.8h\n"
+ "mov v21.16b, v31.16b\n fmla v21.8h, v0.8h, v9.8h\n"
+ "fmla v28.8h, v5.8h, v13.8h\n"
+ "fmla v27.8h, v3.8h, v13.8h\n"
"fmla v26.8h, v2.8h, v13.8h\n"
- "fmla v27.8h, v1.8h, v13.8h\n"
- "fmla v28.8h, v0.8h, v13.8h\n"
- "ldr q13, [x16, x8]\n"
- "fmla v29.8h, v6.8h, v12.8h\n"
- "ldr q12, [x27, x26]\n"
- "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n"
- "ldr q16, [x14, #0x0]\n"
+ "fmla v25.8h, v1.8h, v13.8h\n"
"fmla v24.8h, v0.8h, v13.8h\n"
- "fmla v31.8h, v8.8h, v12.8h\n"
- "ldr q12, [x16, x9]\n"
- "fmla v23.8h, v7.8h, v11.8h\n"
- "fmla v30.8h, v0.8h, v11.8h\n"
- "fmla v26.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v3.8h, v11.8h\n"
- "fmla v29.8h, v1.8h, v11.8h\n"
- "ld1 { v11.8h }, [x13]\n"
- "fmla v24.8h, v2.8h, v12.8h\n"
- "fmla v25.8h, v1.8h, v12.8h\n"
- "ld1 { v12.8h }, [x10]\n"
- "fmla v28.8h, v4.8h, v10.8h\n"
- "fmla v23.8h, v1.8h, v13.8h\n"
- "ldr q13, [x13, x26]\n"
- "fmla v30.8h, v2.8h, v10.8h\n"
- "fmla v31.8h, v1.8h, v10.8h\n"
- "fmla v24.8h, v8.8h, v10.8h\n"
- "fmla v25.8h, v7.8h, v10.8h\n"
- "fmla v27.8h, v5.8h, v10.8h\n"
- "ldr q10, [x10, x11]\n"
- "fmla v26.8h, v0.8h, v11.8h\n"
- "fmla v29.8h, v3.8h, v12.8h\n"
- "fmla v28.8h, v2.8h, v13.8h\n"
- "fmla v30.8h, v4.8h, v10.8h\n"
- "fmla v31.8h, v3.8h, v10.8h\n"
- "fmla v23.8h, v3.8h, v11.8h\n"
- "ldr q11, [x10, x26]\n"
- "fmla v25.8h, v5.8h, v13.8h\n"
- "ldr q13, [x27, x8]\n"
- "fmla v26.8h, v6.8h, v12.8h\n"
- "ldr q12, [x13, x8]\n"
- "fmla v27.8h, v7.8h, v10.8h\n"
- "fmla v29.8h, v5.8h, v10.8h\n"
- "fmla v28.8h, v6.8h, v10.8h\n"
- "fmla v31.8h, v5.8h, v11.8h\n"
- "fmla v30.8h, v6.8h, v13.8h\n"
- "fmla v26.8h, v8.8h, v10.8h\n"
- "fmla v29.8h, v7.8h, v13.8h\n"
- "ldr q13, [x27, x9]\n"
- "fmla v24.8h, v3.8h, v12.8h\n"
- "fmla v27.8h, v0.8h, v12.8h\n"
- "fmla v28.8h, v8.8h, v11.8h\n"
- "ldr q11, [x13, x9]\n"
- "fmla v30.8h, v8.8h, v13.8h\n"
+ "ldr q17, [x16, x8]\n"
+ "fmla v22.8h, v6.8h, v12.8h\n"
+ "ldr q16, [x27, x26]\n"
+ "mov v20.16b, v31.16b\n fmla v20.8h, v1.8h, v9.8h\n"
+ "ldr q31, [x14, #0x0]\n"
+ "fmla v29.8h, v0.8h, v17.8h\n"
+ "fmla v21.8h, v8.8h, v16.8h\n"
+ "ldr q16, [x16, x9]\n"
+ "fmla v28.8h, v7.8h, v18.8h\n"
+ "fmla v20.8h, v0.8h, v18.8h\n"
+ "fmla v26.8h, v4.8h, v18.8h\n"
+ "fmla v25.8h, v3.8h, v18.8h\n"
+ "fmla v22.8h, v1.8h, v18.8h\n"
+ "ld1 { v19.8h }, [x13]\n"
+ "fmla v29.8h, v2.8h, v16.8h\n"
+ "fmla v27.8h, v1.8h, v16.8h\n"
+ "ld1 { v18.8h }, [x10]\n"
+ "fmla v24.8h, v4.8h, v23.8h\n"
+ "fmla v28.8h, v1.8h, v17.8h\n"
+ "ldr q16, [x13, x26]\n"
+ "fmla v20.8h, v2.8h, v23.8h\n"
+ "fmla v21.8h, v1.8h, v23.8h\n"
+ "fmla v29.8h, v8.8h, v23.8h\n"
+ "fmla v27.8h, v7.8h, v23.8h\n"
+ "fmla v25.8h, v5.8h, v23.8h\n"
+ "ldr q17, [x10, x11]\n"
+ "fmla v26.8h, v0.8h, v19.8h\n"
+ "fmla v22.8h, v3.8h, v18.8h\n"
+ "fmla v24.8h, v2.8h, v16.8h\n"
+ "fmla v20.8h, v4.8h, v17.8h\n"
+ "fmla v21.8h, v3.8h, v17.8h\n"
+ "fmla v28.8h, v3.8h, v19.8h\n"
+ "ldr q19, [x10, x26]\n"
+ "fmla v27.8h, v5.8h, v16.8h\n"
+ "ldr q16, [x27, x8]\n"
+ "fmla v26.8h, v6.8h, v18.8h\n"
+ "ldr q18, [x13, x8]\n"
+ "fmla v25.8h, v7.8h, v17.8h\n"
+ "fmla v22.8h, v5.8h, v17.8h\n"
+ "fmla v24.8h, v6.8h, v17.8h\n"
+ "fmla v21.8h, v5.8h, v19.8h\n"
+ "fmla v20.8h, v6.8h, v16.8h\n"
+ "fmla v26.8h, v8.8h, v17.8h\n"
+ "fmla v22.8h, v7.8h, v16.8h\n"
+ "ldr q17, [x27, x9]\n"
+ "fmla v29.8h, v3.8h, v18.8h\n"
+ "fmla v25.8h, v0.8h, v18.8h\n"
+ "fmla v24.8h, v8.8h, v19.8h\n"
+ "ldr q16, [x13, x9]\n"
+ "fmla v20.8h, v8.8h, v17.8h\n"
"add x13, x13, #0x10\n"
- "fmla v31.8h, v7.8h, v13.8h\n"
- "ldr q13, [x10, x9]\n"
- "fmla v23.8h, v4.8h, v12.8h\n"
- "fmla v26.8h, v1.8h, v12.8h\n"
- "ldr q12, [x10, x8]\n"
- "fmla v24.8h, v5.8h, v11.8h\n"
+ "fmla v21.8h, v7.8h, v17.8h\n"
+ "ldr q19, [x10, x9]\n"
+ "fmla v28.8h, v4.8h, v18.8h\n"
+ "fmla v26.8h, v1.8h, v18.8h\n"
+ "ldr q17, [x10, x8]\n"
+ "fmla v29.8h, v5.8h, v16.8h\n"
"add x10, x10, #0x10\n"
- "fmla v25.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v2.8h, v11.8h\n"
- "fmla v28.8h, v1.8h, v11.8h\n"
- "ldr q11, [x16, x11]\n"
- "fmla v29.8h, v4.8h, v12.8h\n"
+ "fmla v27.8h, v4.8h, v16.8h\n"
+ "fmla v25.8h, v2.8h, v16.8h\n"
+ "fmla v24.8h, v1.8h, v16.8h\n"
+ "ldr q16, [x16, x11]\n"
+ "fmla v22.8h, v4.8h, v17.8h\n"
"add x16, x16, #0x10\n"
"ld1 { v10.8h }, [x16]\n"
- "fmla v30.8h, v3.8h, v12.8h\n"
- "fmla v31.8h, v4.8h, v13.8h\n"
+ "fmla v20.8h, v3.8h, v17.8h\n"
+ "fmla v21.8h, v4.8h, v19.8h\n"
"ldr q4, [x14, #0x50]\n"
- "fmla v26.8h, v7.8h, v12.8h\n"
- "fmla v27.8h, v6.8h, v12.8h\n"
- "ld1 { v12.8h }, [x12]\n"
- "fmla v23.8h, v2.8h, v11.8h\n"
- "fmla v24.8h, v1.8h, v11.8h\n"
+ "fmla v26.8h, v7.8h, v17.8h\n"
+ "fmla v25.8h, v6.8h, v17.8h\n"
+ "ld1 { v18.8h }, [x12]\n"
+ "fmla v28.8h, v2.8h, v16.8h\n"
+ "fmla v29.8h, v1.8h, v16.8h\n"
"ldr q1, [x14, #0x20]\n"
- "fmax v24.8h, v24.8h, v18.8h\n"
- "fmla v25.8h, v0.8h, v11.8h\n"
- "ldr q11, [x12, x26]\n"
- "fmla v28.8h, v7.8h, v13.8h\n"
+ "fmax v29.8h, v29.8h, v15.8h\n"
+ "fmla v27.8h, v0.8h, v16.8h\n"
+ "ldr q17, [x12, x26]\n"
+ "fmla v24.8h, v7.8h, v19.8h\n"
"add x12, x12, #0x10\n"
"ldr q9, [x12, x11]\n"
- "fmla v30.8h, v5.8h, v13.8h\n"
- "fmla v29.8h, v0.8h, v12.8h\n"
+ "fmla v20.8h, v5.8h, v19.8h\n"
+ "fmla v22.8h, v0.8h, v18.8h\n"
"ldr q0, [x14, #0x10]\n"
- "fmla v31.8h, v2.8h, v11.8h\n"
+ "fmla v21.8h, v2.8h, v17.8h\n"
"ldr q2, [x14, #0x30]\n"
- "fmla v27.8h, v8.8h, v13.8h\n"
- "ldr q13, [x27, x11]\n"
- "fmla v23.8h, v6.8h, v12.8h\n"
- "fmla v26.8h, v3.8h, v12.8h\n"
+ "fmla v25.8h, v8.8h, v19.8h\n"
+ "ldr q16, [x27, x11]\n"
+ "fmla v28.8h, v6.8h, v18.8h\n"
+ "fmla v26.8h, v3.8h, v18.8h\n"
"ldr q3, [x14, #0x40]\n"
- "fmax v23.8h, v23.8h, v18.8h\n"
- "fmla v25.8h, v8.8h, v11.8h\n"
- "fmla v28.8h, v5.8h, v11.8h\n"
+ "fmax v28.8h, v28.8h, v15.8h\n"
+ "fmla v27.8h, v8.8h, v17.8h\n"
+ "fmla v24.8h, v5.8h, v17.8h\n"
"ldr q11, [x16, x26]\n"
"ldr q5, [x14, #0x60]\n"
- "fmla v29.8h, v8.8h, v13.8h\n"
+ "fmla v22.8h, v8.8h, v16.8h\n"
"ldr q8, [x14, #0x90]\n"
- "fmla v30.8h, v7.8h, v13.8h\n"
+ "fmla v20.8h, v7.8h, v16.8h\n"
"ldr q7, [x14, #0x80]\n"
- "fmla v31.8h, v6.8h, v13.8h\n"
+ "fmla v21.8h, v6.8h, v16.8h\n"
"ldr q13, [x13, x11]\n"
"ldr q6, [x14, #0x70]\n"
- "fmax v25.8h, v25.8h, v18.8h\n"
- "fmax v26.8h, v26.8h, v18.8h\n"
- "fmax v27.8h, v27.8h, v18.8h\n"
+ "fmax v27.8h, v27.8h, v15.8h\n"
+ "fmax v26.8h, v26.8h, v15.8h\n"
+ "fmax v25.8h, v25.8h, v15.8h\n"
"add x27, x27, #0x10\n"
"ld1 { v12.8h }, [x27]\n"
- "fmax v28.8h, v28.8h, v18.8h\n"
- "fmax v29.8h, v29.8h, v18.8h\n"
+ "fmax v24.8h, v24.8h, v15.8h\n"
+ "fmax v22.8h, v22.8h, v15.8h\n"
"add x14, x14, #0xa0\n"
- "fmax v30.8h, v30.8h, v18.8h\n"
- "fmax v31.8h, v31.8h, v18.8h\n"
- "fmin v23.8h, v23.8h, v17.8h\n"
- "fmin v24.8h, v24.8h, v17.8h\n"
- "st1 { v23.8h }, [x15]\n"
- "fmin v25.8h, v25.8h, v17.8h\n"
- "fmin v26.8h, v26.8h, v17.8h\n"
- "str q24, [x15, x17]\n"
- "fmin v27.8h, v27.8h, v17.8h\n"
- "fmin v28.8h, v28.8h, v17.8h\n"
- "str q25, [x15, x22]\n"
+ "fmax v20.8h, v20.8h, v15.8h\n"
+ "fmax v21.8h, v21.8h, v15.8h\n"
+ "fmin v28.8h, v28.8h, v14.8h\n"
+ "fmin v29.8h, v29.8h, v14.8h\n"
+ "st1 { v28.8h }, [x15]\n"
+ "fmin v27.8h, v27.8h, v14.8h\n"
+ "fmin v26.8h, v26.8h, v14.8h\n"
+ "str q29, [x15, x17]\n"
+ "fmin v25.8h, v25.8h, v14.8h\n"
+ "fmin v24.8h, v24.8h, v14.8h\n"
+ "str q27, [x15, x22]\n"
"add x15, x15, #0x10\n"
- "fmin v29.8h, v29.8h, v17.8h\n"
- "fmin v30.8h, v30.8h, v17.8h\n"
+ "fmin v22.8h, v22.8h, v14.8h\n"
+ "fmin v20.8h, v20.8h, v14.8h\n"
"st1 { v26.8h }, [x28]\n"
- "fmin v31.8h, v31.8h, v17.8h\n"
- "str q27, [x28, x17]\n"
- "str q28, [x28, x22]\n"
+ "fmin v21.8h, v21.8h, v14.8h\n"
+ "str q25, [x28, x17]\n"
+ "str q24, [x28, x22]\n"
"add x28, x28, #0x10\n"
- "st1 { v29.8h }, [x25]\n"
- "str q30, [x25, x17]\n"
- "str q31, [x25, x22]\n"
+ "st1 { v22.8h }, [x25]\n"
+ "str q20, [x25, x17]\n"
+ "str q21, [x25, x22]\n"
"add x25, x25, #0x10\n"
"blt 2b\n"
"3:" // Tile loop: Channel tail
- "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n"
- "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n"
- "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n"
- "fmla v24.8h, v4.8h, v13.8h\n"
- "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n"
- "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n"
- "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n"
- "fmla v23.8h, v0.8h, v10.8h\n"
- "ldr q10, [x12, x9]\n"
- "fmla v25.8h, v2.8h, v11.8h\n"
- "ldr q11, [x12, x8]\n"
- "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n"
- "fmla v24.8h, v6.8h, v11.8h\n"
- "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n"
- "fmla v23.8h, v5.8h, v13.8h\n"
- "fmla v25.8h, v3.8h, v13.8h\n"
+ "mov v29.16b, v31.16b\n fmla v29.8h, v7.8h, v9.8h\n"
+ "mov v28.16b, v31.16b\n fmla v28.8h, v8.8h, v9.8h\n"
+ "mov v27.16b, v31.16b\n fmla v27.8h, v6.8h, v9.8h\n"
+ "fmla v29.8h, v4.8h, v13.8h\n"
+ "mov v26.16b, v31.16b\n fmla v26.8h, v5.8h, v9.8h\n"
+ "mov v25.16b, v31.16b\n fmla v25.8h, v4.8h, v9.8h\n"
+ "mov v24.16b, v31.16b\n fmla v24.8h, v3.8h, v9.8h\n"
+ "fmla v28.8h, v0.8h, v10.8h\n"
+ "ldr q23, [x12, x9]\n"
+ "fmla v27.8h, v2.8h, v11.8h\n"
+ "ldr q18, [x12, x8]\n"
+ "mov v22.16b, v31.16b\n fmla v22.8h, v2.8h, v9.8h\n"
+ "fmla v29.8h, v6.8h, v18.8h\n"
+ "mov v21.16b, v31.16b\n fmla v21.8h, v0.8h, v9.8h\n"
+ "fmla v28.8h, v5.8h, v13.8h\n"
+ "fmla v27.8h, v3.8h, v13.8h\n"
"fmla v26.8h, v2.8h, v13.8h\n"
- "fmla v27.8h, v1.8h, v13.8h\n"
- "fmla v28.8h, v0.8h, v13.8h\n"
- "ldr q13, [x16, x8]\n"
- "fmla v29.8h, v6.8h, v12.8h\n"
- "ldr q12, [x27, x26]\n"
- "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n"
+ "fmla v25.8h, v1.8h, v13.8h\n"
"fmla v24.8h, v0.8h, v13.8h\n"
- "fmla v31.8h, v8.8h, v12.8h\n"
- "ldr q12, [x16, x9]\n"
- "fmla v23.8h, v7.8h, v11.8h\n"
- "fmla v30.8h, v0.8h, v11.8h\n"
- "fmla v26.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v3.8h, v11.8h\n"
- "fmla v29.8h, v1.8h, v11.8h\n"
- "ld1 { v11.8h }, [x13]\n"
- "fmla v24.8h, v2.8h, v12.8h\n"
- "fmla v25.8h, v1.8h, v12.8h\n"
- "ld1 { v12.8h }, [x10]\n"
- "fmla v28.8h, v4.8h, v10.8h\n"
- "fmla v23.8h, v1.8h, v13.8h\n"
- "ldr q13, [x13, x26]\n"
- "fmla v30.8h, v2.8h, v10.8h\n"
- "fmla v31.8h, v1.8h, v10.8h\n"
- "fmla v24.8h, v8.8h, v10.8h\n"
- "fmla v25.8h, v7.8h, v10.8h\n"
- "fmla v27.8h, v5.8h, v10.8h\n"
- "ldr q10, [x10, x11]\n"
- "fmla v26.8h, v0.8h, v11.8h\n"
- "fmla v29.8h, v3.8h, v12.8h\n"
- "fmla v28.8h, v2.8h, v13.8h\n"
- "fmla v30.8h, v4.8h, v10.8h\n"
- "fmla v31.8h, v3.8h, v10.8h\n"
- "fmla v23.8h, v3.8h, v11.8h\n"
- "ldr q11, [x10, x26]\n"
- "fmla v25.8h, v5.8h, v13.8h\n"
- "ldr q13, [x27, x8]\n"
- "fmla v26.8h, v6.8h, v12.8h\n"
- "ldr q12, [x13, x8]\n"
- "fmla v27.8h, v7.8h, v10.8h\n"
- "fmla v29.8h, v5.8h, v10.8h\n"
- "fmla v28.8h, v6.8h, v10.8h\n"
- "fmla v31.8h, v5.8h, v11.8h\n"
- "fmla v30.8h, v6.8h, v13.8h\n"
- "fmla v26.8h, v8.8h, v10.8h\n"
- "fmla v29.8h, v7.8h, v13.8h\n"
- "ldr q13, [x27, x9]\n"
- "fmla v24.8h, v3.8h, v12.8h\n"
- "fmla v27.8h, v0.8h, v12.8h\n"
- "fmla v28.8h, v8.8h, v11.8h\n"
- "ldr q11, [x13, x9]\n"
- "fmla v30.8h, v8.8h, v13.8h\n"
+ "ldr q17, [x16, x8]\n"
+ "fmla v22.8h, v6.8h, v12.8h\n"
+ "ldr q16, [x27, x26]\n"
+ "mov v20.16b, v31.16b\n fmla v20.8h, v1.8h, v9.8h\n"
+ "fmla v29.8h, v0.8h, v17.8h\n"
+ "fmla v21.8h, v8.8h, v16.8h\n"
+ "ldr q16, [x16, x9]\n"
+ "fmla v28.8h, v7.8h, v18.8h\n"
+ "fmla v20.8h, v0.8h, v18.8h\n"
+ "fmla v26.8h, v4.8h, v18.8h\n"
+ "fmla v25.8h, v3.8h, v18.8h\n"
+ "fmla v22.8h, v1.8h, v18.8h\n"
+ "ld1 { v19.8h }, [x13]\n"
+ "fmla v29.8h, v2.8h, v16.8h\n"
+ "fmla v27.8h, v1.8h, v16.8h\n"
+ "ld1 { v18.8h }, [x10]\n"
+ "fmla v24.8h, v4.8h, v23.8h\n"
+ "fmla v28.8h, v1.8h, v17.8h\n"
+ "ldr q16, [x13, x26]\n"
+ "fmla v20.8h, v2.8h, v23.8h\n"
+ "fmla v21.8h, v1.8h, v23.8h\n"
+ "fmla v29.8h, v8.8h, v23.8h\n"
+ "fmla v27.8h, v7.8h, v23.8h\n"
+ "fmla v25.8h, v5.8h, v23.8h\n"
+ "ldr q17, [x10, x11]\n"
+ "fmla v26.8h, v0.8h, v19.8h\n"
+ "fmla v22.8h, v3.8h, v18.8h\n"
+ "fmla v24.8h, v2.8h, v16.8h\n"
+ "fmla v20.8h, v4.8h, v17.8h\n"
+ "fmla v21.8h, v3.8h, v17.8h\n"
+ "fmla v28.8h, v3.8h, v19.8h\n"
+ "ldr q19, [x10, x26]\n"
+ "fmla v27.8h, v5.8h, v16.8h\n"
+ "ldr q16, [x27, x8]\n"
+ "fmla v26.8h, v6.8h, v18.8h\n"
+ "ldr q18, [x13, x8]\n"
+ "fmla v25.8h, v7.8h, v17.8h\n"
+ "fmla v22.8h, v5.8h, v17.8h\n"
+ "fmla v24.8h, v6.8h, v17.8h\n"
+ "fmla v21.8h, v5.8h, v19.8h\n"
+ "fmla v20.8h, v6.8h, v16.8h\n"
+ "fmla v26.8h, v8.8h, v17.8h\n"
+ "fmla v22.8h, v7.8h, v16.8h\n"
+ "ldr q17, [x27, x9]\n"
+ "fmla v29.8h, v3.8h, v18.8h\n"
+ "fmla v25.8h, v0.8h, v18.8h\n"
+ "fmla v24.8h, v8.8h, v19.8h\n"
+ "ldr q16, [x13, x9]\n"
+ "fmla v20.8h, v8.8h, v17.8h\n"
"add x13, x13, #0x10\n"
- "fmla v31.8h, v7.8h, v13.8h\n"
- "ldr q13, [x10, x9]\n"
- "fmla v23.8h, v4.8h, v12.8h\n"
- "fmla v26.8h, v1.8h, v12.8h\n"
- "ldr q12, [x10, x8]\n"
- "fmla v24.8h, v5.8h, v11.8h\n"
+ "fmla v21.8h, v7.8h, v17.8h\n"
+ "ldr q19, [x10, x9]\n"
+ "fmla v28.8h, v4.8h, v18.8h\n"
+ "fmla v26.8h, v1.8h, v18.8h\n"
+ "ldr q17, [x10, x8]\n"
+ "fmla v29.8h, v5.8h, v16.8h\n"
"add x10, x10, #0x10\n"
- "fmla v25.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v2.8h, v11.8h\n"
- "fmla v28.8h, v1.8h, v11.8h\n"
- "ldr q11, [x16, x11]\n"
- "fmla v29.8h, v4.8h, v12.8h\n"
+ "fmla v27.8h, v4.8h, v16.8h\n"
+ "fmla v25.8h, v2.8h, v16.8h\n"
+ "fmla v24.8h, v1.8h, v16.8h\n"
+ "ldr q16, [x16, x11]\n"
+ "fmla v22.8h, v4.8h, v17.8h\n"
"add x16, x16, #0x10\n"
- "fmla v30.8h, v3.8h, v12.8h\n"
- "fmla v31.8h, v4.8h, v13.8h\n"
- "fmla v26.8h, v7.8h, v12.8h\n"
- "fmla v27.8h, v6.8h, v12.8h\n"
- "ld1 { v12.8h }, [x12]\n"
- "fmla v23.8h, v2.8h, v11.8h\n"
- "fmla v24.8h, v1.8h, v11.8h\n"
- "fmax v24.8h, v24.8h, v18.8h\n"
- "fmla v25.8h, v0.8h, v11.8h\n"
- "ldr q11, [x12, x26]\n"
- "fmla v28.8h, v7.8h, v13.8h\n"
- "fmin v24.8h, v24.8h, v17.8h\n"
- "fmla v30.8h, v5.8h, v13.8h\n"
- "fmla v29.8h, v0.8h, v12.8h\n"
+ "fmla v20.8h, v3.8h, v17.8h\n"
+ "fmla v21.8h, v4.8h, v19.8h\n"
+ "fmla v26.8h, v7.8h, v17.8h\n"
+ "fmla v25.8h, v6.8h, v17.8h\n"
+ "ld1 { v18.8h }, [x12]\n"
+ "fmla v28.8h, v2.8h, v16.8h\n"
+ "fmla v29.8h, v1.8h, v16.8h\n"
+ "fmax v29.8h, v29.8h, v15.8h\n"
+ "fmla v27.8h, v0.8h, v16.8h\n"
+ "ldr q17, [x12, x26]\n"
+ "fmla v24.8h, v7.8h, v19.8h\n"
+ "fmin v29.8h, v29.8h, v14.8h\n"
+ "fmla v20.8h, v5.8h, v19.8h\n"
+ "fmla v22.8h, v0.8h, v18.8h\n"
"add x12, x12, #0x10\n"
- "fmla v31.8h, v2.8h, v11.8h\n"
- "fmla v27.8h, v8.8h, v13.8h\n"
- "ldr q13, [x27, x11]\n"
- "fmax v27.8h, v27.8h, v18.8h\n"
- "fmla v23.8h, v6.8h, v12.8h\n"
- "fmla v26.8h, v3.8h, v12.8h\n"
- "fmax v23.8h, v23.8h, v18.8h\n"
+ "fmla v21.8h, v2.8h, v17.8h\n"
+ "fmla v25.8h, v8.8h, v19.8h\n"
+ "ldr q16, [x27, x11]\n"
+ "fmax v25.8h, v25.8h, v15.8h\n"
+ "fmla v28.8h, v6.8h, v18.8h\n"
+ "fmla v26.8h, v3.8h, v18.8h\n"
+ "fmax v28.8h, v28.8h, v15.8h\n"
"add x27, x27, #0x10\n"
- "fmla v25.8h, v8.8h, v11.8h\n"
- "fmla v28.8h, v5.8h, v11.8h\n"
- "fmax v25.8h, v25.8h, v18.8h\n"
- "fmla v29.8h, v8.8h, v13.8h\n"
- "fmla v30.8h, v7.8h, v13.8h\n"
- "fmax v26.8h, v26.8h, v18.8h\n"
- "fmla v31.8h, v6.8h, v13.8h\n"
- "fmax v28.8h, v28.8h, v18.8h\n"
- "fmax v29.8h, v29.8h, v18.8h\n"
- "fmax v30.8h, v30.8h, v18.8h\n"
- "fmax v31.8h, v31.8h, v18.8h\n"
- "fmin v23.8h, v23.8h, v17.8h\n"
- "st1 { v23.8h }, [x15]\n"
- "fmin v25.8h, v25.8h, v17.8h\n"
- "fmin v26.8h, v26.8h, v17.8h\n"
- "str q24, [x15, x17]\n"
- "fmin v27.8h, v27.8h, v17.8h\n"
- "fmin v28.8h, v28.8h, v17.8h\n"
- "str q25, [x15, x22]\n"
+ "fmla v27.8h, v8.8h, v17.8h\n"
+ "fmla v24.8h, v5.8h, v17.8h\n"
+ "fmax v27.8h, v27.8h, v15.8h\n"
+ "fmla v22.8h, v8.8h, v16.8h\n"
+ "fmla v20.8h, v7.8h, v16.8h\n"
+ "fmax v26.8h, v26.8h, v15.8h\n"
+ "fmla v21.8h, v6.8h, v16.8h\n"
+ "fmax v24.8h, v24.8h, v15.8h\n"
+ "fmax v22.8h, v22.8h, v15.8h\n"
+ "fmax v20.8h, v20.8h, v15.8h\n"
+ "fmax v21.8h, v21.8h, v15.8h\n"
+ "fmin v28.8h, v28.8h, v14.8h\n"
+ "st1 { v28.8h }, [x15]\n"
+ "fmin v27.8h, v27.8h, v14.8h\n"
+ "fmin v26.8h, v26.8h, v14.8h\n"
+ "str q29, [x15, x17]\n"
+ "fmin v25.8h, v25.8h, v14.8h\n"
+ "fmin v24.8h, v24.8h, v14.8h\n"
+ "str q27, [x15, x22]\n"
"add x15, x15, #0x10\n"
- "fmin v29.8h, v29.8h, v17.8h\n"
- "fmin v30.8h, v30.8h, v17.8h\n"
+ "fmin v22.8h, v22.8h, v14.8h\n"
+ "fmin v20.8h, v20.8h, v14.8h\n"
"st1 { v26.8h }, [x28]\n"
- "fmin v31.8h, v31.8h, v17.8h\n"
- "str q27, [x28, x17]\n"
- "str q28, [x28, x22]\n"
+ "fmin v21.8h, v21.8h, v14.8h\n"
+ "str q25, [x28, x17]\n"
+ "str q24, [x28, x22]\n"
"add x28, x28, #0x10\n"
- "st1 { v29.8h }, [x25]\n"
- "str q30, [x25, x17]\n"
- "str q31, [x25, x22]\n"
+ "st1 { v22.8h }, [x25]\n"
+ "str q20, [x25, x17]\n"
+ "str q21, [x25, x22]\n"
"add x25, x25, #0x10\n"
"4:" // Tile loop: Oddments
"tst %x[n_channels], #0x7\n"
"beq 93f\n"
- "ldr q16, [x14, #0x0]\n"
+ "ldr q31, [x14, #0x0]\n"
"ldr q0, [x14, #0x10]\n"
"add x24, x12, x11\n"
"add x23, x16, XZR\n"
@@ -509,18 +509,18 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"ldr h12, [x21, #0x0]\n"
"ldr h13, [x20, #0x0]\n"
"8:" // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 2: End
- "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n"
- "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n"
+ "mov v23.16b, v31.16b\n fmla v23.8h, v8.8h, v9.8h\n"
+ "mov v25.16b, v31.16b\n fmla v25.8h, v6.8h, v9.8h\n"
"add x20, x27, x26\n"
- "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n"
- "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n"
- "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n"
- "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n"
- "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n"
+ "mov v24.16b, v31.16b\n fmla v24.8h, v7.8h, v9.8h\n"
+ "mov v26.16b, v31.16b\n fmla v26.8h, v5.8h, v9.8h\n"
+ "mov v27.16b, v31.16b\n fmla v27.8h, v4.8h, v9.8h\n"
+ "mov v28.16b, v31.16b\n fmla v28.8h, v3.8h, v9.8h\n"
+ "mov v29.16b, v31.16b\n fmla v29.8h, v2.8h, v9.8h\n"
"fmla v23.8h, v0.8h, v10.8h\n"
"fmla v25.8h, v2.8h, v11.8h\n"
- "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n"
- "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n"
+ "mov v30.16b, v31.16b\n fmla v30.8h, v1.8h, v9.8h\n"
+ "fmla v31.8h, v0.8h, v9.8h\n"
"fmla v29.8h, v6.8h, v12.8h\n"
"fmla v23.8h, v5.8h, v13.8h\n"
"fmla v24.8h, v4.8h, v13.8h\n"
@@ -1009,25 +1009,25 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"88:" // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: End
"fmla v29.8h, v8.8h, v13.8h\n"
"fmla v30.8h, v7.8h, v13.8h\n"
- "fmax v23.8h, v23.8h, v18.8h\n"
+ "fmax v23.8h, v23.8h, v15.8h\n"
"fmla v31.8h, v6.8h, v13.8h\n"
- "fmax v24.8h, v24.8h, v18.8h\n"
- "fmax v25.8h, v25.8h, v18.8h\n"
- "fmax v26.8h, v26.8h, v18.8h\n"
- "fmax v27.8h, v27.8h, v18.8h\n"
- "fmax v28.8h, v28.8h, v18.8h\n"
- "fmax v29.8h, v29.8h, v18.8h\n"
- "fmax v30.8h, v30.8h, v18.8h\n"
- "fmax v31.8h, v31.8h, v18.8h\n"
- "fmin v23.8h, v23.8h, v17.8h\n"
- "fmin v24.8h, v24.8h, v17.8h\n"
- "fmin v25.8h, v25.8h, v17.8h\n"
- "fmin v26.8h, v26.8h, v17.8h\n"
- "fmin v27.8h, v27.8h, v17.8h\n"
- "fmin v28.8h, v28.8h, v17.8h\n"
- "fmin v29.8h, v29.8h, v17.8h\n"
- "fmin v30.8h, v30.8h, v17.8h\n"
- "fmin v31.8h, v31.8h, v17.8h\n"
+ "fmax v24.8h, v24.8h, v15.8h\n"
+ "fmax v25.8h, v25.8h, v15.8h\n"
+ "fmax v26.8h, v26.8h, v15.8h\n"
+ "fmax v27.8h, v27.8h, v15.8h\n"
+ "fmax v28.8h, v28.8h, v15.8h\n"
+ "fmax v29.8h, v29.8h, v15.8h\n"
+ "fmax v30.8h, v30.8h, v15.8h\n"
+ "fmax v31.8h, v31.8h, v15.8h\n"
+ "fmin v23.8h, v23.8h, v14.8h\n"
+ "fmin v24.8h, v24.8h, v14.8h\n"
+ "fmin v25.8h, v25.8h, v14.8h\n"
+ "fmin v26.8h, v26.8h, v14.8h\n"
+ "fmin v27.8h, v27.8h, v14.8h\n"
+ "fmin v28.8h, v28.8h, v14.8h\n"
+ "fmin v29.8h, v29.8h, v14.8h\n"
+ "fmin v30.8h, v30.8h, v14.8h\n"
+ "fmin v31.8h, v31.8h, v14.8h\n"
"tbz %x[n_channels], #2, 90f\n"
"mov x22, x15\n"
"mov x21, x28\n"
@@ -1134,7 +1134,6 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl(
"st1 { v28.h }[0], [x21]\n"
"st1 { v31.h }[0], [x20]\n"
"92:" // Tile loop: Oddments: Store: Bit 2: End
-
"93:" // Tile loop: End
"ldr x23, [%x[params_struct], %[offsetof_args_tile_j]]\n"
"ldr x24, [%x[params_struct], %[offsetof_args_tile_i]]\n"
@@ -1149,7 +1148,7 @@ void a64_fp16_nhwc_3x3_s1_output3x3_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", "v16", "v17", "v18", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "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", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}