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 | 515 |
1 files changed, 258 insertions, 257 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 7eea14f70f..f4927c5536 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 @@ -22,12 +22,12 @@ * SOFTWARE. */ -#if defined(__aarch64__) - #include "pooling.hpp" #include <cstdint> #include <cstddef> +#if defined(__aarch64__) + namespace arm_conv { namespace pooling { @@ -43,77 +43,77 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( { __asm__ __volatile__( "cmp %x[n_channels], #0x40\n" - "mov x9, #0x0\n" - "mov x28, #0x10\n" // cntb _, ALL, #1 - "mov x27, #0x20\n" // cntb _, ALL, #2 - "mov x26, #0x30\n" // cntb _, ALL, #3 + "mov x27, #0x0\n" + "mov x26, #0x10\n" // cntb _, ALL, #1 + "mov x24, #0x20\n" // cntb _, ALL, #2 + "mov x23, #0x30\n" // cntb _, ALL, #3 "blt 7f\n" "1:" // 4-vectors of channels "lsr x25, %x[n_valid_cells], #0x2\n" "movi v8.16b, #0x0\n" "movi v7.16b, #0x0\n" - "mov x20, %x[inptrs]\n" + "mov x22, %x[inptrs]\n" "movi v6.16b, #0x0\n" "movi v5.16b, #0x0\n" "cbz x25, 4f\n" - "ldp x24, x23, [x20, #0x0]\n" - "ldr q4, [x24, x9]\n" + "ldp x21, x20, [x22, #0x0]\n" + "ldr q4, [x21, x27]\n" "subs x25, x25, #0x1\n" - "ldr q3, [x23, x9]\n" - "ldr q2, [x24, x28]\n" - "ldr q1, [x23, x28]\n" - "ldr q0, [x24, x27]\n" - "ldr q31, [x23, x27]\n" - "ldr q30, [x24, x26]\n" - "ldr q29, [x23, x26]\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" - "ldr q28, [x22, x9]\n" - "ldr q22, [x21, x9]\n" - "ldr q27, [x22, x28]\n" - "ldr q21, [x21, x28]\n" - "ldr q26, [x22, x27]\n" - "ldr q20, [x21, x27]\n" - "ldr q25, [x22, x26]\n" - "ldr q24, [x21, x26]\n" + "ldr q3, [x20, x27]\n" + "ldr q2, [x21, x26]\n" + "ldr q1, [x20, x26]\n" + "ldr q0, [x21, x24]\n" + "ldr q31, [x20, x24]\n" + "ldr q30, [x21, x23]\n" + "ldr q29, [x20, x23]\n" + "ldp x21, x20, [x22, #0x10]\n" + "add x22, x22, #0x20\n" + "ldr q28, [x21, x27]\n" + "ldr q22, [x20, x27]\n" + "ldr q27, [x21, x26]\n" + "ldr q21, [x20, x26]\n" + "ldr q26, [x21, x24]\n" + "ldr q20, [x20, x24]\n" + "ldr q25, [x21, x23]\n" + "ldr q24, [x20, x23]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "umax v23.16b, v4.16b, v3.16b\n" "umax v19.16b, v28.16b, v22.16b\n" - "ldp x24, x23, [x20, #0x0]\n" - "ldr q4, [x24, x9]\n" - "ldr q3, [x23, x9]\n" + "ldp x21, x20, [x22, #0x0]\n" + "ldr q4, [x21, x27]\n" + "ldr q3, [x20, x27]\n" "umax v22.16b, v2.16b, v1.16b\n" - "ldr q2, [x24, x28]\n" + "ldr q2, [x21, x26]\n" "umax v18.16b, v27.16b, v21.16b\n" - "ldr q1, [x23, x28]\n" + "ldr q1, [x20, x26]\n" "umax v21.16b, v0.16b, v31.16b\n" - "ldr q0, [x24, x27]\n" + "ldr q0, [x21, x24]\n" "umax v17.16b, v26.16b, v20.16b\n" - "ldr q31, [x23, x27]\n" + "ldr q31, [x20, x24]\n" "umax v20.16b, v30.16b, v29.16b\n" - "ldr q30, [x24, x26]\n" + "ldr q30, [x21, x23]\n" "umax v16.16b, v25.16b, v24.16b\n" - "ldr q29, [x23, x26]\n" + "ldr q29, [x20, x23]\n" "umax v19.16b, v23.16b, v19.16b\n" "umax v18.16b, v22.16b, v18.16b\n" - "ldp x22, x21, [x20, #0x10]\n" - "ldr q28, [x22, x9]\n" - "ldr q22, [x21, x9]\n" + "ldp x21, x20, [x22, #0x10]\n" + "ldr q28, [x21, x27]\n" + "ldr q22, [x20, x27]\n" "umax v17.16b, v21.16b, v17.16b\n" "umax v16.16b, v20.16b, v16.16b\n" - "ldr q27, [x22, x28]\n" - "ldr q21, [x21, x28]\n" + "ldr q27, [x21, x26]\n" + "ldr q21, [x20, x26]\n" "subs x25, x25, #0x1\n" "umax v8.16b, v8.16b, v19.16b\n" - "ldr q26, [x22, x27]\n" - "ldr q20, [x21, x27]\n" + "ldr q26, [x21, x24]\n" + "ldr q20, [x20, x24]\n" "umax v7.16b, v7.16b, v18.16b\n" "umax v6.16b, v6.16b, v17.16b\n" - "ldr q25, [x22, x26]\n" - "ldr q24, [x21, x26]\n" + "ldr q25, [x21, x23]\n" + "ldr q24, [x20, x23]\n" "umax v5.16b, v5.16b, v16.16b\n" - "add x20, x20, #0x20\n" + "add x22, x22, #0x20\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "umax v23.16b, v4.16b, v3.16b\n" @@ -136,16 +136,16 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x24, [x20], #0x8\n" - "ldr q4, [x24, x9]\n" + "ldr x20, [x22], #0x8\n" + "ldr q16, [x20, x27]\n" "subs x21, x21, #0x1\n" - "umax v8.16b, v8.16b, v4.16b\n" - "ldr q2, [x24, x28]\n" - "ldr q0, [x24, x27]\n" - "umax v7.16b, v7.16b, v2.16b\n" - "umax v6.16b, v6.16b, v0.16b\n" - "ldr q30, [x24, x26]\n" - "umax v5.16b, v5.16b, v30.16b\n" + "umax v8.16b, v8.16b, v16.16b\n" + "ldr q17, [x20, x26]\n" + "ldr q16, [x20, x24]\n" + "umax v7.16b, v7.16b, v17.16b\n" + "umax v6.16b, v6.16b, v16.16b\n" + "ldr q16, [x20, x23]\n" + "umax v5.16b, v5.16b, v16.16b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End "add x20, %x[quant_params], %[offsetof_qp_input_offset]\n" @@ -292,17 +292,17 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "uzp1 v19.16b, v25.16b, v19.16b\n" "uzp1 v18.16b, v24.16b, v18.16b\n" "uzp1 v16.16b, v23.16b, v16.16b\n" - "str q16, [%x[outptr], x9]\n" - "add x9, x9, #0x40\n" + "str q16, [%x[outptr], x27]\n" + "add x27, x27, #0x40\n" "uzp1 v16.16b, v22.16b, v21.16b\n" "uzp1 v17.16b, v20.16b, v17.16b\n" - "str q16, [%x[outptr], x28]\n" - "add x28, x28, #0x40\n" - "uzp1 v16.16b, v19.16b, v18.16b\n" - "str q17, [%x[outptr], x27]\n" - "add x27, x27, #0x40\n" "str q16, [%x[outptr], x26]\n" "add x26, x26, #0x40\n" + "uzp1 v16.16b, v19.16b, v18.16b\n" + "str q17, [%x[outptr], x24]\n" + "add x24, x24, #0x40\n" + "str q16, [%x[outptr], x23]\n" + "add x23, x23, #0x40\n" "bge 1b\n" "cbz %x[n_channels], 43f\n" "7:" // Single vector of channels @@ -311,314 +311,314 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" "movi v8.16b, #0x0\n" - "mov x20, %x[inptrs]\n" + "mov x22, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x24, x23, [x20, #0x0]\n" - "ldr q4, [x24, x9]\n" + "ldp x21, x20, [x22, #0x0]\n" + "ldr q4, [x21, x27]\n" "subs x25, x25, #0x1\n" - "ldr q3, [x23, x9]\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" - "ldr q28, [x22, x9]\n" - "ldr q22, [x21, x9]\n" + "ldr q3, [x20, x27]\n" + "ldp x21, x20, [x22, #0x10]\n" + "add x22, x22, #0x20\n" + "ldr q28, [x21, x27]\n" + "ldr q22, [x20, x27]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop - "umax v23.16b, v4.16b, v3.16b\n" - "umax v19.16b, v28.16b, v22.16b\n" - "ldp x24, x23, [x20, #0x0]\n" - "ldr q4, [x24, x9]\n" - "ldr q3, [x23, x9]\n" - "umax v19.16b, v23.16b, v19.16b\n" - "ldp x22, x21, [x20, #0x10]\n" + "umax v17.16b, v4.16b, v3.16b\n" + "umax v16.16b, v28.16b, v22.16b\n" + "ldp x21, x20, [x22, #0x0]\n" + "ldr q4, [x21, x27]\n" + "ldr q3, [x20, x27]\n" + "umax v16.16b, v17.16b, v16.16b\n" + "ldp x21, x20, [x22, #0x10]\n" "subs x25, x25, #0x1\n" - "ldr q28, [x22, x9]\n" - "ldr q22, [x21, x9]\n" - "umax v8.16b, v8.16b, v19.16b\n" - "add x20, x20, #0x20\n" + "ldr q28, [x21, x27]\n" + "ldr q22, [x20, x27]\n" + "umax v8.16b, v8.16b, v16.16b\n" + "add x22, x22, #0x20\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail - "umax v23.16b, v4.16b, v3.16b\n" - "umax v19.16b, v28.16b, v22.16b\n" - "umax v19.16b, v23.16b, v19.16b\n" - "umax v8.16b, v8.16b, v19.16b\n" + "umax v17.16b, v4.16b, v3.16b\n" + "umax v16.16b, v28.16b, v22.16b\n" + "umax v16.16b, v17.16b, v16.16b\n" + "umax v8.16b, v8.16b, v16.16b\n" "11:" // Single vector of channels: Loop: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x24, [x20], #0x8\n" - "ldr q4, [x24, x9]\n" + "ldr x20, [x22], #0x8\n" + "ldr q16, [x20, x27]\n" "subs x21, x21, #0x1\n" - "umax v8.16b, v8.16b, v4.16b\n" + "umax v8.16b, v8.16b, v16.16b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End "add x20, %x[quant_params], %[offsetof_qp_input_offset]\n" - "ld1r { v4.4s }, [x20]\n" - "uxtl v23.8h, v8.8b\n" - "uxtl2 v24.8h, v8.16b\n" - "neg v4.4s, v4.4s\n" + "ld1r { v18.4s }, [x20]\n" + "uxtl v17.8h, v8.8b\n" + "uxtl2 v16.8h, v8.16b\n" + "neg v18.4s, v18.4s\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1r { v3.4s }, [x20]\n" - "saddw v0.4s, v4.4s, v23.4h\n" - "saddw2 v23.4s, v4.4s, v23.8h\n" - "saddw v31.4s, v4.4s, v24.4h\n" + "ld1r { v23.4s }, [x20]\n" + "saddw v22.4s, v18.4s, v17.4h\n" + "saddw2 v21.4s, v18.4s, v17.8h\n" + "saddw v20.4s, v18.4s, v16.4h\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "ld1r { v2.4s }, [x20]\n" - "saddw2 v30.4s, v4.4s, v24.8h\n" - "srshl v0.4s, v0.4s, v3.4s\n" + "ld1r { v19.4s }, [x20]\n" + "saddw2 v18.4s, v18.4s, v16.8h\n" + "srshl v22.4s, v22.4s, v23.4s\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1r { v1.4s }, [x20]\n" - "srshl v23.4s, v23.4s, v3.4s\n" - "srshl v31.4s, v31.4s, v3.4s\n" + "ld1r { v17.4s }, [x20]\n" + "srshl v21.4s, v21.4s, v23.4s\n" + "srshl v20.4s, v20.4s, v23.4s\n" "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" "ld1r { v16.4s }, [x20]\n" - "srshl v30.4s, v30.4s, v3.4s\n" - "sqrdmulh v0.4s, v0.4s, v2.4s\n" + "srshl v18.4s, v18.4s, v23.4s\n" + "sqrdmulh v22.4s, v22.4s, v19.4s\n" "sub %x[n_channels], %x[n_channels], #0x10\n" "cmp %x[n_channels], #0x10\n" - "sqrdmulh v23.4s, v23.4s, v2.4s\n" - "sqrdmulh v31.4s, v31.4s, v2.4s\n" - "sqrdmulh v30.4s, v30.4s, v2.4s\n" - "srshl v0.4s, v0.4s, v1.4s\n" - "srshl v23.4s, v23.4s, v1.4s\n" - "srshl v31.4s, v31.4s, v1.4s\n" - "srshl v30.4s, v30.4s, v1.4s\n" - "add v0.4s, v0.4s, v16.4s\n" - "add v23.4s, v23.4s, v16.4s\n" - "add v31.4s, v31.4s, v16.4s\n" - "add v30.4s, v30.4s, v16.4s\n" + "sqrdmulh v21.4s, v21.4s, v19.4s\n" + "sqrdmulh v20.4s, v20.4s, v19.4s\n" + "sqrdmulh v18.4s, v18.4s, v19.4s\n" + "srshl v22.4s, v22.4s, v17.4s\n" + "srshl v21.4s, v21.4s, v17.4s\n" + "srshl v20.4s, v20.4s, v17.4s\n" + "srshl v18.4s, v18.4s, v17.4s\n" + "add v22.4s, v22.4s, v16.4s\n" + "add v21.4s, v21.4s, v16.4s\n" + "add v20.4s, v20.4s, v16.4s\n" + "add v18.4s, v18.4s, v16.4s\n" "movi v16.4s, #0x0\n" - "smax v0.4s, v0.4s, v16.4s\n" - "smax v23.4s, v23.4s, v16.4s\n" - "smax v31.4s, v31.4s, v16.4s\n" - "smax v30.4s, v30.4s, v16.4s\n" + "smax v22.4s, v22.4s, v16.4s\n" + "smax v21.4s, v21.4s, v16.4s\n" + "smax v20.4s, v20.4s, v16.4s\n" + "smax v18.4s, v18.4s, v16.4s\n" "movi v16.4s, #0xff\n" - "smin v0.4s, v0.4s, v16.4s\n" - "smin v23.4s, v23.4s, v16.4s\n" - "smin v31.4s, v31.4s, v16.4s\n" - "smin v30.4s, v30.4s, v16.4s\n" - "uzp1 v23.16b, v0.16b, v23.16b\n" - "uzp1 v16.16b, v31.16b, v30.16b\n" - "uzp1 v16.16b, v23.16b, v16.16b\n" - "str q16, [%x[outptr], x9]\n" - "add x9, x9, #0x10\n" + "smin v22.4s, v22.4s, v16.4s\n" + "smin v21.4s, v21.4s, v16.4s\n" + "smin v20.4s, v20.4s, v16.4s\n" + "smin v18.4s, v18.4s, v16.4s\n" + "uzp1 v17.16b, v22.16b, v21.16b\n" + "uzp1 v16.16b, v20.16b, v18.16b\n" + "uzp1 v16.16b, v17.16b, v16.16b\n" + "str q16, [%x[outptr], x27]\n" + "add x27, x27, #0x10\n" "bge 8b\n" "cbz %x[n_channels], 43f\n" "14:" // Oddments "lsr x25, %x[n_valid_cells], #0x2\n" - "add %x[outptr], %x[outptr], x9\n" + "add %x[outptr], %x[outptr], x27\n" "movi v8.16b, #0x0\n" - "mov x20, %x[inptrs]\n" + "mov x24, %x[inptrs]\n" "cbz x25, 24f\n" "15:" // Oddments: 4 inputs loop - "ldp x24, x23, [x20, #0x0]\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" - "add x24, x24, x9\n" - "add x23, x23, x9\n" - "add x22, x22, x9\n" + "ldp x23, x22, [x24, #0x0]\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "add x23, x23, x27\n" + "add x22, x22, x27\n" + "add x21, x21, x27\n" "movi v4.16b, #0x0\n" "movi v3.16b, #0x0\n" - "add x21, x21, x9\n" + "add x20, x20, x27\n" "movi v28.16b, #0x0\n" "movi v22.16b, #0x0\n" "tbz %x[n_channels], #3, 19f\n" - "ldr d4, [x24], #0x8\n" - "ldr d3, [x23], #0x8\n" - "ldr d28, [x22], #0x8\n" - "ldr d22, [x21], #0x8\n" + "ldr d4, [x23], #0x8\n" + "ldr d3, [x22], #0x8\n" + "ldr d28, [x21], #0x8\n" + "ldr d22, [x20], #0x8\n" "tbz %x[n_channels], #2, 17f\n" - "ld1 { v4.s }[2], [x24], #0x4\n" - "ld1 { v3.s }[2], [x23], #0x4\n" - "ld1 { v28.s }[2], [x22], #0x4\n" - "ld1 { v22.s }[2], [x21], #0x4\n" + "ld1 { v4.s }[2], [x23], #0x4\n" + "ld1 { v3.s }[2], [x22], #0x4\n" + "ld1 { v28.s }[2], [x21], #0x4\n" + "ld1 { v22.s }[2], [x20], #0x4\n" "tbz %x[n_channels], #1, 16f\n" - "ld1 { v4.h }[6], [x24], #0x2\n" - "ld1 { v3.h }[6], [x23], #0x2\n" - "ld1 { v28.h }[6], [x22], #0x2\n" - "ld1 { v22.h }[6], [x21], #0x2\n" + "ld1 { v4.h }[6], [x23], #0x2\n" + "ld1 { v3.h }[6], [x22], #0x2\n" + "ld1 { v28.h }[6], [x21], #0x2\n" + "ld1 { v22.h }[6], [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v4.b }[14], [x24], #0x1\n" - "ld1 { v3.b }[14], [x23], #0x1\n" - "ld1 { v28.b }[14], [x22], #0x1\n" - "ld1 { v22.b }[14], [x21], #0x1\n" + "ld1 { v4.b }[14], [x23], #0x1\n" + "ld1 { v3.b }[14], [x22], #0x1\n" + "ld1 { v28.b }[14], [x21], #0x1\n" + "ld1 { v22.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 { v4.b }[12], [x24], #0x1\n" - "ld1 { v3.b }[12], [x23], #0x1\n" - "ld1 { v28.b }[12], [x22], #0x1\n" - "ld1 { v22.b }[12], [x21], #0x1\n" + "ld1 { v4.b }[12], [x23], #0x1\n" + "ld1 { v3.b }[12], [x22], #0x1\n" + "ld1 { v28.b }[12], [x21], #0x1\n" + "ld1 { v22.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 { v4.h }[4], [x24], #0x2\n" - "ld1 { v3.h }[4], [x23], #0x2\n" - "ld1 { v28.h }[4], [x22], #0x2\n" - "ld1 { v22.h }[4], [x21], #0x2\n" + "ld1 { v4.h }[4], [x23], #0x2\n" + "ld1 { v3.h }[4], [x22], #0x2\n" + "ld1 { v28.h }[4], [x21], #0x2\n" + "ld1 { v22.h }[4], [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v4.b }[10], [x24], #0x1\n" - "ld1 { v3.b }[10], [x23], #0x1\n" - "ld1 { v28.b }[10], [x22], #0x1\n" - "ld1 { v22.b }[10], [x21], #0x1\n" + "ld1 { v4.b }[10], [x23], #0x1\n" + "ld1 { v3.b }[10], [x22], #0x1\n" + "ld1 { v28.b }[10], [x21], #0x1\n" + "ld1 { v22.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 { v4.b }[8], [x24], #0x1\n" - "ld1 { v3.b }[8], [x23], #0x1\n" - "ld1 { v28.b }[8], [x22], #0x1\n" - "ld1 { v22.b }[8], [x21], #0x1\n" + "ld1 { v4.b }[8], [x23], #0x1\n" + "ld1 { v3.b }[8], [x22], #0x1\n" + "ld1 { v28.b }[8], [x21], #0x1\n" + "ld1 { v22.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 s4, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr s28, [x22], #0x4\n" - "ldr s22, [x21], #0x4\n" + "ldr s4, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s28, [x21], #0x4\n" + "ldr s22, [x20], #0x4\n" "tbz %x[n_channels], #1, 20f\n" - "ld1 { v4.h }[2], [x24], #0x2\n" - "ld1 { v3.h }[2], [x23], #0x2\n" - "ld1 { v28.h }[2], [x22], #0x2\n" - "ld1 { v22.h }[2], [x21], #0x2\n" + "ld1 { v4.h }[2], [x23], #0x2\n" + "ld1 { v3.h }[2], [x22], #0x2\n" + "ld1 { v28.h }[2], [x21], #0x2\n" + "ld1 { v22.h }[2], [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v4.b }[6], [x24], #0x1\n" - "ld1 { v3.b }[6], [x23], #0x1\n" - "ld1 { v28.b }[6], [x22], #0x1\n" - "ld1 { v22.b }[6], [x21], #0x1\n" + "ld1 { v4.b }[6], [x23], #0x1\n" + "ld1 { v3.b }[6], [x22], #0x1\n" + "ld1 { v28.b }[6], [x21], #0x1\n" + "ld1 { v22.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 { v4.b }[4], [x24], #0x1\n" - "ld1 { v3.b }[4], [x23], #0x1\n" - "ld1 { v28.b }[4], [x22], #0x1\n" - "ld1 { v22.b }[4], [x21], #0x1\n" + "ld1 { v4.b }[4], [x23], #0x1\n" + "ld1 { v3.b }[4], [x22], #0x1\n" + "ld1 { v28.b }[4], [x21], #0x1\n" + "ld1 { v22.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 h4, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" - "ldr h28, [x22], #0x2\n" - "ldr h22, [x21], #0x2\n" + "ldr h4, [x23], #0x2\n" + "ldr h3, [x22], #0x2\n" + "ldr h28, [x21], #0x2\n" + "ldr h22, [x20], #0x2\n" "tbz %x[n_channels], #0, 23f\n" - "ld1 { v4.b }[2], [x24], #0x1\n" - "ld1 { v3.b }[2], [x23], #0x1\n" - "ld1 { v28.b }[2], [x22], #0x1\n" - "ld1 { v22.b }[2], [x21], #0x1\n" + "ld1 { v4.b }[2], [x23], #0x1\n" + "ld1 { v3.b }[2], [x22], #0x1\n" + "ld1 { v28.b }[2], [x21], #0x1\n" + "ld1 { v22.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 b4, [x24], #0x1\n" - "ldr b3, [x23], #0x1\n" - "ldr b28, [x22], #0x1\n" - "ldr b22, [x21], #0x1\n" + "ldr b4, [x23], #0x1\n" + "ldr b3, [x22], #0x1\n" + "ldr b28, [x21], #0x1\n" + "ldr b22, [x20], #0x1\n" "23:" // Oddments: 4 inputs loop: Load: Bit 3: End - "umax v23.16b, v4.16b, v3.16b\n" - "umax v19.16b, v28.16b, v22.16b\n" + "umax v17.16b, v4.16b, v3.16b\n" + "umax v16.16b, v28.16b, v22.16b\n" "subs x25, x25, #0x1\n" - "umax v19.16b, v23.16b, v19.16b\n" - "umax v8.16b, v8.16b, v19.16b\n" + "umax v16.16b, v17.16b, v16.16b\n" + "umax v8.16b, v8.16b, v16.16b\n" "bgt 15b\n" "24:" // Oddments: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 34f\n" "25:" // Oddments: Single input loop - "ldr x24, [x20], #0x8\n" - "add x24, x24, x9\n" + "ldr x23, [x24], #0x8\n" + "add x23, x23, x27\n" "movi v4.16b, #0x0\n" "tbz %x[n_channels], #3, 29f\n" - "ldr d4, [x24], #0x8\n" + "ldr d4, [x23], #0x8\n" "tbz %x[n_channels], #2, 27f\n" - "ld1 { v4.s }[2], [x24], #0x4\n" + "ld1 { v4.s }[2], [x23], #0x4\n" "tbz %x[n_channels], #1, 26f\n" - "ld1 { v4.h }[6], [x24], #0x2\n" + "ld1 { v4.h }[6], [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v4.b }[14], [x24], #0x1\n" + "ld1 { v4.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 { v4.b }[12], [x24], #0x1\n" + "ld1 { v4.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 { v4.h }[4], [x24], #0x2\n" + "ld1 { v4.h }[4], [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v4.b }[10], [x24], #0x1\n" + "ld1 { v4.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 { v4.b }[8], [x24], #0x1\n" + "ld1 { v4.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 s4, [x24], #0x4\n" + "ldr s4, [x23], #0x4\n" "tbz %x[n_channels], #1, 30f\n" - "ld1 { v4.h }[2], [x24], #0x2\n" + "ld1 { v4.h }[2], [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v4.b }[6], [x24], #0x1\n" + "ld1 { v4.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 { v4.b }[4], [x24], #0x1\n" + "ld1 { v4.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 h4, [x24], #0x2\n" + "ldr h4, [x23], #0x2\n" "tbz %x[n_channels], #0, 33f\n" - "ld1 { v4.b }[2], [x24], #0x1\n" + "ld1 { v4.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 b4, [x24], #0x1\n" + "ldr b4, [x23], #0x1\n" "33:" // Oddments: Single input loop: Load: Bit 3: End "subs x21, x21, #0x1\n" "umax v8.16b, v8.16b, v4.16b\n" "bgt 25b\n" "34:" // Oddments: Single input loop: End "add x20, %x[quant_params], %[offsetof_qp_input_offset]\n" - "ld1r { v4.4s }, [x20]\n" - "uxtl v23.8h, v8.8b\n" - "uxtl2 v24.8h, v8.16b\n" - "neg v4.4s, v4.4s\n" + "ld1r { v18.4s }, [x20]\n" + "uxtl v17.8h, v8.8b\n" + "uxtl2 v16.8h, v8.16b\n" + "neg v18.4s, v18.4s\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "ld1r { v3.4s }, [x20]\n" - "saddw v0.4s, v4.4s, v23.4h\n" - "saddw2 v23.4s, v4.4s, v23.8h\n" - "saddw v31.4s, v4.4s, v24.4h\n" + "ld1r { v23.4s }, [x20]\n" + "saddw v22.4s, v18.4s, v17.4h\n" + "saddw2 v21.4s, v18.4s, v17.8h\n" + "saddw v20.4s, v18.4s, v16.4h\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - "ld1r { v2.4s }, [x20]\n" - "saddw2 v30.4s, v4.4s, v24.8h\n" - "srshl v0.4s, v0.4s, v3.4s\n" + "ld1r { v19.4s }, [x20]\n" + "saddw2 v18.4s, v18.4s, v16.8h\n" + "srshl v22.4s, v22.4s, v23.4s\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "ld1r { v1.4s }, [x20]\n" - "srshl v23.4s, v23.4s, v3.4s\n" - "srshl v31.4s, v31.4s, v3.4s\n" + "ld1r { v17.4s }, [x20]\n" + "srshl v21.4s, v21.4s, v23.4s\n" + "srshl v20.4s, v20.4s, v23.4s\n" "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" "ld1r { v16.4s }, [x20]\n" - "srshl v30.4s, v30.4s, v3.4s\n" - "sqrdmulh v0.4s, v0.4s, v2.4s\n" - "sqrdmulh v23.4s, v23.4s, v2.4s\n" - "sqrdmulh v31.4s, v31.4s, v2.4s\n" - "sqrdmulh v30.4s, v30.4s, v2.4s\n" - "srshl v0.4s, v0.4s, v1.4s\n" - "srshl v23.4s, v23.4s, v1.4s\n" - "srshl v31.4s, v31.4s, v1.4s\n" - "srshl v30.4s, v30.4s, v1.4s\n" - "add v0.4s, v0.4s, v16.4s\n" - "add v23.4s, v23.4s, v16.4s\n" - "add v31.4s, v31.4s, v16.4s\n" - "add v30.4s, v30.4s, v16.4s\n" + "srshl v18.4s, v18.4s, v23.4s\n" + "sqrdmulh v22.4s, v22.4s, v19.4s\n" + "sqrdmulh v21.4s, v21.4s, v19.4s\n" + "sqrdmulh v20.4s, v20.4s, v19.4s\n" + "sqrdmulh v18.4s, v18.4s, v19.4s\n" + "srshl v22.4s, v22.4s, v17.4s\n" + "srshl v21.4s, v21.4s, v17.4s\n" + "srshl v20.4s, v20.4s, v17.4s\n" + "srshl v18.4s, v18.4s, v17.4s\n" + "add v22.4s, v22.4s, v16.4s\n" + "add v21.4s, v21.4s, v16.4s\n" + "add v20.4s, v20.4s, v16.4s\n" + "add v18.4s, v18.4s, v16.4s\n" "movi v16.4s, #0x0\n" - "smax v0.4s, v0.4s, v16.4s\n" - "smax v23.4s, v23.4s, v16.4s\n" - "smax v31.4s, v31.4s, v16.4s\n" - "smax v30.4s, v30.4s, v16.4s\n" + "smax v22.4s, v22.4s, v16.4s\n" + "smax v21.4s, v21.4s, v16.4s\n" + "smax v20.4s, v20.4s, v16.4s\n" + "smax v18.4s, v18.4s, v16.4s\n" "movi v16.4s, #0xff\n" - "smin v0.4s, v0.4s, v16.4s\n" - "smin v23.4s, v23.4s, v16.4s\n" - "smin v31.4s, v31.4s, v16.4s\n" - "smin v30.4s, v30.4s, v16.4s\n" - "uzp1 v23.16b, v0.16b, v23.16b\n" - "uzp1 v16.16b, v31.16b, v30.16b\n" - "uzp1 v16.16b, v23.16b, v16.16b\n" + "smin v22.4s, v22.4s, v16.4s\n" + "smin v21.4s, v21.4s, v16.4s\n" + "smin v20.4s, v20.4s, v16.4s\n" + "smin v18.4s, v18.4s, v16.4s\n" + "uzp1 v17.16b, v22.16b, v21.16b\n" + "uzp1 v16.16b, v20.16b, v18.16b\n" + "uzp1 v16.16b, v17.16b, v16.16b\n" "tbz %x[n_channels], #3, 38f\n" "st1 { v16.d }[0], [%x[outptr]], #0x8\n" "tbz %x[n_channels], #2, 36f\n" @@ -667,10 +667,11 @@ void a64_u8q_nhwc_max_generic_depthfirst_impl( "43:" // End : [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", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "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", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27" ); } } // namespace pooling } // namespace arm_conv -#endif // defined(__aarch64__) + +#endif // defined(__aarch64__) |