diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp | 270 |
1 files changed, 135 insertions, 135 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp index b056be20bc..0d196e097e 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp @@ -51,7 +51,7 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "1:" // 4-vectors of channels "movi v4.16b, #0x0\n" "mov x19, %x[inptrs]\n" - "movi v3.16b, #0x0\n" + "movi v8.16b, #0x0\n" "lsr x24, %x[n_valid_cells], #0x2\n" "movi v7.16b, #0x0\n" "movi v6.16b, #0x0\n" @@ -60,12 +60,12 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "ldp x21, x20, [x19, #0x10]\n" "add x19, x19, #0x20\n" "subs x24, x24, #0x1\n" - "ldr q2, [x23, x28]\n" - "ldr q1, [x22, x28]\n" - "ldr q0, [x21, x28]\n" - "ldr q31, [x20, x28]\n" - "ldr q30, [x23, x27]\n" - "ldr q22, [x22, x27]\n" + "ldr q3, [x23, x28]\n" + "ldr q2, [x22, x28]\n" + "ldr q1, [x21, x28]\n" + "ldr q0, [x20, x28]\n" + "ldr q31, [x23, x27]\n" + "ldr q30, [x22, x27]\n" "ldr q29, [x21, x27]\n" "ldr q28, [x20, x27]\n" "ldr q27, [x23, x26]\n" @@ -78,47 +78,47 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "ldr q16, [x20, x25]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop - "umax v23.16b, v2.16b, v1.16b\n" + "umax v23.16b, v3.16b, v2.16b\n" "ldp x23, x22, [x19, #0x0]\n" "subs x24, x24, #0x1\n" - "umax v19.16b, v0.16b, v31.16b\n" + "umax v19.16b, v1.16b, v0.16b\n" "ldp x21, x20, [x19, #0x10]\n" - "umax v22.16b, v30.16b, v22.16b\n" "add x19, x19, #0x20\n" + "umax v22.16b, v31.16b, v30.16b\n" + "ldr q3, [x23, x28]\n" "umax v18.16b, v29.16b, v28.16b\n" - "ldr q2, [x23, x28]\n" "umax v21.16b, v27.16b, v21.16b\n" + "ldr q2, [x22, x28]\n" "umax v17.16b, v26.16b, v17.16b\n" - "ldr q1, [x22, x28]\n" + "ldr q1, [x21, x28]\n" "umax v20.16b, v25.16b, v20.16b\n" - "ldr q0, [x21, x28]\n" + "ldr q0, [x20, x28]\n" "umax v16.16b, v24.16b, v16.16b\n" - "ldr q31, [x20, x28]\n" + "ldr q31, [x23, x27]\n" "umax v19.16b, v23.16b, v19.16b\n" - "ldr q30, [x23, x27]\n" + "ldr q30, [x22, x27]\n" "umax v18.16b, v22.16b, v18.16b\n" - "ldr q22, [x22, x27]\n" - "umax v17.16b, v21.16b, v17.16b\n" "ldr q29, [x21, x27]\n" - "umax v16.16b, v20.16b, v16.16b\n" + "umax v17.16b, v21.16b, v17.16b\n" "ldr q28, [x20, x27]\n" - "umax v4.16b, v4.16b, v19.16b\n" + "umax v16.16b, v20.16b, v16.16b\n" "ldr q27, [x23, x26]\n" - "umax v3.16b, v3.16b, v18.16b\n" + "umax v4.16b, v4.16b, v19.16b\n" "ldr q21, [x22, x26]\n" - "umax v7.16b, v7.16b, v17.16b\n" + "umax v8.16b, v8.16b, v18.16b\n" "ldr q26, [x21, x26]\n" - "umax v6.16b, v6.16b, v16.16b\n" + "umax v7.16b, v7.16b, v17.16b\n" "ldr q17, [x20, x26]\n" + "umax v6.16b, v6.16b, v16.16b\n" "ldr q25, [x23, x25]\n" "ldr q20, [x22, x25]\n" "ldr q24, [x21, x25]\n" "ldr q16, [x20, x25]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail - "umax v23.16b, v2.16b, v1.16b\n" - "umax v19.16b, v0.16b, v31.16b\n" - "umax v22.16b, v30.16b, v22.16b\n" + "umax v23.16b, v3.16b, v2.16b\n" + "umax v19.16b, v1.16b, v0.16b\n" + "umax v22.16b, v31.16b, v30.16b\n" "umax v18.16b, v29.16b, v28.16b\n" "umax v21.16b, v27.16b, v21.16b\n" "umax v17.16b, v26.16b, v17.16b\n" @@ -129,7 +129,7 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "umax v17.16b, v21.16b, v17.16b\n" "umax v16.16b, v20.16b, v16.16b\n" "umax v4.16b, v4.16b, v19.16b\n" - "umax v3.16b, v3.16b, v18.16b\n" + "umax v8.16b, v8.16b, v18.16b\n" "umax v7.16b, v7.16b, v17.16b\n" "umax v6.16b, v6.16b, v16.16b\n" "4:" // 4-vectors of channels: After loop @@ -138,11 +138,11 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "5:" // 4-vectors of channels: Single input loop "ldr x23, [x19], #0x8\n" "subs x20, x20, #0x1\n" - "ldr q2, [x23, x28]\n" - "umax v4.16b, v4.16b, v2.16b\n" - "ldr q30, [x23, x27]\n" + "ldr q3, [x23, x28]\n" + "umax v4.16b, v4.16b, v3.16b\n" + "ldr q31, [x23, x27]\n" "ldr q27, [x23, x26]\n" - "umax v3.16b, v3.16b, v30.16b\n" + "umax v8.16b, v8.16b, v31.16b\n" "ldr q25, [x23, x25]\n" "umax v7.16b, v7.16b, v27.16b\n" "umax v6.16b, v6.16b, v25.16b\n" @@ -153,10 +153,10 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "ld1r { v5.4s }, [x19]\n" "uxtl2 v16.8h, v4.16b\n" "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "uxtl v21.8h, v3.8b\n" + "uxtl v21.8h, v8.8b\n" "ld1r { v4.4s }, [x19]\n" "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "uxtl2 v20.8h, v3.16b\n" + "uxtl2 v20.8h, v8.16b\n" "ld1r { v3.4s }, [x19]\n" "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" "uxtl v19.8h, v7.8b\n" @@ -287,18 +287,18 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "uzp1 v16.16b, v26.16b, v24.16b\n" "str q16, [%x[outptr], x28]\n" "add v22.4s, v22.4s, v1.4s\n" - "add v21.4s, v21.4s, v1.4s\n" "add x28, x28, #0x40\n" + "add v21.4s, v21.4s, v1.4s\n" "uzp1 v16.16b, v23.16b, v20.16b\n" "str q16, [%x[outptr], x27]\n" "smax v22.4s, v22.4s, v30.4s\n" - "smax v21.4s, v21.4s, v30.4s\n" "add x27, x27, #0x40\n" + "smax v21.4s, v21.4s, v30.4s\n" "uzp1 v16.16b, v19.16b, v18.16b\n" "str q16, [%x[outptr], x26]\n" "smin v22.4s, v22.4s, v29.4s\n" - "smin v21.4s, v21.4s, v29.4s\n" "add x26, x26, #0x40\n" + "smin v21.4s, v21.4s, v29.4s\n" "uzp1 v16.16b, v22.16b, v21.16b\n" "uzp1 v16.16b, v17.16b, v16.16b\n" "str q16, [%x[outptr], x25]\n" @@ -314,31 +314,31 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "lsr x24, %x[n_valid_cells], #0x2\n" "cbz x24, 11f\n" "ldp x23, x22, [x19, #0x0]\n" - "subs x24, x24, #0x1\n" "ldp x21, x20, [x19, #0x10]\n" "add x19, x19, #0x20\n" - "ldr q2, [x23, x28]\n" - "ldr q1, [x22, x28]\n" - "ldr q0, [x21, x28]\n" - "ldr q31, [x20, x28]\n" + "subs x24, x24, #0x1\n" + "ldr q3, [x23, x28]\n" + "ldr q2, [x22, x28]\n" + "ldr q1, [x21, x28]\n" + "ldr q0, [x20, x28]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop - "umax v23.16b, v2.16b, v1.16b\n" + "umax v23.16b, v3.16b, v2.16b\n" "ldp x23, x22, [x19, #0x0]\n" "subs x24, x24, #0x1\n" - "umax v19.16b, v0.16b, v31.16b\n" + "umax v19.16b, v1.16b, v0.16b\n" "ldp x21, x20, [x19, #0x10]\n" - "umax v19.16b, v23.16b, v19.16b\n" "add x19, x19, #0x20\n" - "ldr q2, [x23, x28]\n" + "umax v19.16b, v23.16b, v19.16b\n" + "ldr q3, [x23, x28]\n" + "ldr q2, [x22, x28]\n" "umax v4.16b, v4.16b, v19.16b\n" - "ldr q1, [x22, x28]\n" - "ldr q0, [x21, x28]\n" - "ldr q31, [x20, x28]\n" + "ldr q1, [x21, x28]\n" + "ldr q0, [x20, x28]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail - "umax v23.16b, v2.16b, v1.16b\n" - "umax v19.16b, v0.16b, v31.16b\n" + "umax v23.16b, v3.16b, v2.16b\n" + "umax v19.16b, v1.16b, v0.16b\n" "umax v19.16b, v23.16b, v19.16b\n" "umax v4.16b, v4.16b, v19.16b\n" "11:" // Single vector of channels: Loop: After loop @@ -347,8 +347,8 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "12:" // Single vector of channels: Loop: Single input loop "ldr x23, [x19], #0x8\n" "subs x20, x20, #0x1\n" - "ldr q2, [x23, x28]\n" - "umax v4.16b, v4.16b, v2.16b\n" + "ldr q3, [x23, x28]\n" + "umax v4.16b, v4.16b, v3.16b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End "uxtl v17.8h, v4.8b\n" @@ -410,110 +410,110 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "lsr x24, %x[n_valid_cells], #0x2\n" "cbz x24, 24f\n" "15:" // Oddments: 4 inputs loop - "movi v2.16b, #0x0\n" + "movi v3.16b, #0x0\n" "ldp x23, x22, [x19, #0x0]\n" "add x23, x23, x28\n" - "movi v1.16b, #0x0\n" + "movi v2.16b, #0x0\n" "ldp x21, x20, [x19, #0x10]\n" - "movi v0.16b, #0x0\n" + "movi v1.16b, #0x0\n" "add x19, x19, #0x20\n" - "movi v31.16b, #0x0\n" + "movi v0.16b, #0x0\n" "add x22, x22, x28\n" "add x21, x21, x28\n" "add x20, x20, x28\n" "tbz %x[n_channels], #3, 19f\n" - "ldr d2, [x23], #0x8\n" - "ldr d1, [x22], #0x8\n" - "ldr d0, [x21], #0x8\n" - "ldr d31, [x20], #0x8\n" + "ldr d3, [x23], #0x8\n" + "ldr d2, [x22], #0x8\n" + "ldr d1, [x21], #0x8\n" + "ldr d0, [x20], #0x8\n" "tbz %x[n_channels], #2, 17f\n" - "ld1 { v2.s }[2], [x23], #0x4\n" - "ld1 { v1.s }[2], [x22], #0x4\n" - "ld1 { v0.s }[2], [x21], #0x4\n" - "ld1 { v31.s }[2], [x20], #0x4\n" + "ld1 { v3.s }[2], [x23], #0x4\n" + "ld1 { v2.s }[2], [x22], #0x4\n" + "ld1 { v1.s }[2], [x21], #0x4\n" + "ld1 { v0.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #1, 16f\n" - "ld1 { v2.h }[6], [x23], #0x2\n" - "ld1 { v1.h }[6], [x22], #0x2\n" - "ld1 { v0.h }[6], [x21], #0x2\n" - "ld1 { v31.h }[6], [x20], #0x2\n" + "ld1 { v3.h }[6], [x23], #0x2\n" + "ld1 { v2.h }[6], [x22], #0x2\n" + "ld1 { v1.h }[6], [x21], #0x2\n" + "ld1 { v0.h }[6], [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v2.b }[14], [x23], #0x1\n" - "ld1 { v1.b }[14], [x22], #0x1\n" - "ld1 { v0.b }[14], [x21], #0x1\n" - "ld1 { v31.b }[14], [x20], #0x1\n" + "ld1 { v3.b }[14], [x23], #0x1\n" + "ld1 { v2.b }[14], [x22], #0x1\n" + "ld1 { v1.b }[14], [x21], #0x1\n" + "ld1 { v0.b }[14], [x20], #0x1\n" "b 23f\n" "16:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ld1 { v2.b }[12], [x23], #0x1\n" - "ld1 { v1.b }[12], [x22], #0x1\n" - "ld1 { v0.b }[12], [x21], #0x1\n" - "ld1 { v31.b }[12], [x20], #0x1\n" + "ld1 { v3.b }[12], [x23], #0x1\n" + "ld1 { v2.b }[12], [x22], #0x1\n" + "ld1 { v1.b }[12], [x21], #0x1\n" + "ld1 { v0.b }[12], [x20], #0x1\n" "b 23f\n" "17:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset "tbz %x[n_channels], #1, 18f\n" - "ld1 { v2.h }[4], [x23], #0x2\n" - "ld1 { v1.h }[4], [x22], #0x2\n" - "ld1 { v0.h }[4], [x21], #0x2\n" - "ld1 { v31.h }[4], [x20], #0x2\n" + "ld1 { v3.h }[4], [x23], #0x2\n" + "ld1 { v2.h }[4], [x22], #0x2\n" + "ld1 { v1.h }[4], [x21], #0x2\n" + "ld1 { v0.h }[4], [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v2.b }[10], [x23], #0x1\n" - "ld1 { v1.b }[10], [x22], #0x1\n" - "ld1 { v0.b }[10], [x21], #0x1\n" - "ld1 { v31.b }[10], [x20], #0x1\n" + "ld1 { v3.b }[10], [x23], #0x1\n" + "ld1 { v2.b }[10], [x22], #0x1\n" + "ld1 { v1.b }[10], [x21], #0x1\n" + "ld1 { v0.b }[10], [x20], #0x1\n" "b 23f\n" "18:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ld1 { v2.b }[8], [x23], #0x1\n" - "ld1 { v1.b }[8], [x22], #0x1\n" - "ld1 { v0.b }[8], [x21], #0x1\n" - "ld1 { v31.b }[8], [x20], #0x1\n" + "ld1 { v3.b }[8], [x23], #0x1\n" + "ld1 { v2.b }[8], [x22], #0x1\n" + "ld1 { v1.b }[8], [x21], #0x1\n" + "ld1 { v0.b }[8], [x20], #0x1\n" "b 23f\n" "19:" // Oddments: 4 inputs loop: Load: Bit 3: Unset "tbz %x[n_channels], #2, 21f\n" - "ldr s2, [x23], #0x4\n" - "ldr s1, [x22], #0x4\n" - "ldr s0, [x21], #0x4\n" - "ldr s31, [x20], #0x4\n" + "ldr s3, [x23], #0x4\n" + "ldr s2, [x22], #0x4\n" + "ldr s1, [x21], #0x4\n" + "ldr s0, [x20], #0x4\n" "tbz %x[n_channels], #1, 20f\n" - "ld1 { v2.h }[2], [x23], #0x2\n" - "ld1 { v1.h }[2], [x22], #0x2\n" - "ld1 { v0.h }[2], [x21], #0x2\n" - "ld1 { v31.h }[2], [x20], #0x2\n" + "ld1 { v3.h }[2], [x23], #0x2\n" + "ld1 { v2.h }[2], [x22], #0x2\n" + "ld1 { v1.h }[2], [x21], #0x2\n" + "ld1 { v0.h }[2], [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v2.b }[6], [x23], #0x1\n" - "ld1 { v1.b }[6], [x22], #0x1\n" - "ld1 { v0.b }[6], [x21], #0x1\n" - "ld1 { v31.b }[6], [x20], #0x1\n" + "ld1 { v3.b }[6], [x23], #0x1\n" + "ld1 { v2.b }[6], [x22], #0x1\n" + "ld1 { v1.b }[6], [x21], #0x1\n" + "ld1 { v0.b }[6], [x20], #0x1\n" "b 23f\n" "20:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ld1 { v2.b }[4], [x23], #0x1\n" - "ld1 { v1.b }[4], [x22], #0x1\n" - "ld1 { v0.b }[4], [x21], #0x1\n" - "ld1 { v31.b }[4], [x20], #0x1\n" + "ld1 { v3.b }[4], [x23], #0x1\n" + "ld1 { v2.b }[4], [x22], #0x1\n" + "ld1 { v1.b }[4], [x21], #0x1\n" + "ld1 { v0.b }[4], [x20], #0x1\n" "b 23f\n" "21:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset "tbz %x[n_channels], #1, 22f\n" - "ldr h2, [x23], #0x2\n" - "ldr h1, [x22], #0x2\n" - "ldr h0, [x21], #0x2\n" - "ldr h31, [x20], #0x2\n" + "ldr h3, [x23], #0x2\n" + "ldr h2, [x22], #0x2\n" + "ldr h1, [x21], #0x2\n" + "ldr h0, [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v2.b }[2], [x23], #0x1\n" - "ld1 { v1.b }[2], [x22], #0x1\n" - "ld1 { v0.b }[2], [x21], #0x1\n" - "ld1 { v31.b }[2], [x20], #0x1\n" + "ld1 { v3.b }[2], [x23], #0x1\n" + "ld1 { v2.b }[2], [x22], #0x1\n" + "ld1 { v1.b }[2], [x21], #0x1\n" + "ld1 { v0.b }[2], [x20], #0x1\n" "b 23f\n" "22:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 23f\n" - "ldr b2, [x23], #0x1\n" - "ldr b1, [x22], #0x1\n" - "ldr b0, [x21], #0x1\n" - "ldr b31, [x20], #0x1\n" + "ldr b3, [x23], #0x1\n" + "ldr b2, [x22], #0x1\n" + "ldr b1, [x21], #0x1\n" + "ldr b0, [x20], #0x1\n" "23:" // Oddments: 4 inputs loop: Load: Bit 3: End - "umax v23.16b, v2.16b, v1.16b\n" + "umax v23.16b, v3.16b, v2.16b\n" "subs x24, x24, #0x1\n" - "umax v19.16b, v0.16b, v31.16b\n" + "umax v19.16b, v1.16b, v0.16b\n" "umax v19.16b, v23.16b, v19.16b\n" "umax v4.16b, v4.16b, v19.16b\n" "bgt 15b\n" @@ -521,55 +521,55 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "ands x20, %x[n_valid_cells], #0x3\n" "beq 34f\n" "25:" // Oddments: Single input loop - "movi v2.16b, #0x0\n" + "movi v3.16b, #0x0\n" "ldr x23, [x19], #0x8\n" "add x23, x23, x28\n" "tbz %x[n_channels], #3, 29f\n" - "ldr d2, [x23], #0x8\n" + "ldr d3, [x23], #0x8\n" "tbz %x[n_channels], #2, 27f\n" - "ld1 { v2.s }[2], [x23], #0x4\n" + "ld1 { v3.s }[2], [x23], #0x4\n" "tbz %x[n_channels], #1, 26f\n" - "ld1 { v2.h }[6], [x23], #0x2\n" + "ld1 { v3.h }[6], [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v2.b }[14], [x23], #0x1\n" + "ld1 { v3.b }[14], [x23], #0x1\n" "b 33f\n" "26:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ld1 { v2.b }[12], [x23], #0x1\n" + "ld1 { v3.b }[12], [x23], #0x1\n" "b 33f\n" "27:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset "tbz %x[n_channels], #1, 28f\n" - "ld1 { v2.h }[4], [x23], #0x2\n" + "ld1 { v3.h }[4], [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v2.b }[10], [x23], #0x1\n" + "ld1 { v3.b }[10], [x23], #0x1\n" "b 33f\n" "28:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ld1 { v2.b }[8], [x23], #0x1\n" + "ld1 { v3.b }[8], [x23], #0x1\n" "b 33f\n" "29:" // Oddments: Single input loop: Load: Bit 3: Unset "tbz %x[n_channels], #2, 31f\n" - "ldr s2, [x23], #0x4\n" + "ldr s3, [x23], #0x4\n" "tbz %x[n_channels], #1, 30f\n" - "ld1 { v2.h }[2], [x23], #0x2\n" + "ld1 { v3.h }[2], [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v2.b }[6], [x23], #0x1\n" + "ld1 { v3.b }[6], [x23], #0x1\n" "b 33f\n" "30:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ld1 { v2.b }[4], [x23], #0x1\n" + "ld1 { v3.b }[4], [x23], #0x1\n" "b 33f\n" "31:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset "tbz %x[n_channels], #1, 32f\n" - "ldr h2, [x23], #0x2\n" + "ldr h3, [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v2.b }[2], [x23], #0x1\n" + "ld1 { v3.b }[2], [x23], #0x1\n" "b 33f\n" "32:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset "tbz %x[n_channels], #0, 33f\n" - "ldr b2, [x23], #0x1\n" + "ldr b3, [x23], #0x1\n" "33:" // Oddments: Single input loop: Load: Bit 3: End - "umax v4.16b, v4.16b, v2.16b\n" + "umax v4.16b, v4.16b, v3.16b\n" "subs x20, x20, #0x1\n" "bgt 25b\n" "34:" // Oddments: Single input loop: End @@ -669,7 +669,7 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( : [n_channels] "+&r" (n_channels), [outptr] "+&r" (outptr) : [inptrs] "r" (inptrs), [n_valid_cells] "r" (n_valid_cells), [offsetof_qp_input_offset] "I" (offsetof(Requantize32, input_offset)), [offsetof_qp_output_offset] "I" (offsetof(Requantize32, output_offset)), [offsetof_qp_per_layer_left_shift] "I" (offsetof(Requantize32, per_layer_left_shift)), [offsetof_qp_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_qp_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [quant_params] "r" (&qp) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |