diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp index 797a8f9235..94bf38c29b 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2024 Arm Limited. + * Copyright (c) 2021-2024, 2026 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,20 +43,20 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( { __asm__ __volatile__( "cmp %x[n_channels], #0x40\n" - "mov x9, #0x0\n" + "mov x9, #0\n" "mov x28, #0x10\n" // cntb _, ALL, #1 "mov x27, #0x20\n" // cntb _, ALL, #2 "mov x26, #0x30\n" // cntb _, ALL, #3 "blt 7f\n" "1:" // 4-vectors of channels "lsr x25, %x[n_valid_cells], #0x2\n" + "movi v6.16b, #0x80\n" "movi v9.16b, #0x80\n" - "movi v8.16b, #0x80\n" "mov x24, %x[inptrs]\n" + "movi v8.16b, #0x80\n" "movi v7.16b, #0x80\n" - "movi v6.16b, #0x80\n" "cbz x25, 4f\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "subs x25, x25, #0x1\n" "add x24, x24, #0x20\n" @@ -80,7 +80,7 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "2:" // 4-vectors of channels: 4 inputs loop "smax v23.16b, v5.16b, v4.16b\n" "smax v19.16b, v3.16b, v2.16b\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "smax v22.16b, v1.16b, v0.16b\n" "smax v18.16b, v31.16b, v30.16b\n" @@ -106,14 +106,14 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "smax v16.16b, v20.16b, v16.16b\n" "ldr q28, [x21, x27]\n" "ldr q27, [x20, x27]\n" - "smax v9.16b, v9.16b, v19.16b\n" - "smax v8.16b, v8.16b, v18.16b\n" + "smax v6.16b, v6.16b, v19.16b\n" + "smax v9.16b, v9.16b, v18.16b\n" "ldr q26, [x23, x26]\n" "ldr q20, [x22, x26]\n" - "smax v7.16b, v7.16b, v17.16b\n" + "smax v8.16b, v8.16b, v17.16b\n" "ldr q25, [x21, x26]\n" "ldr q24, [x20, x26]\n" - "smax v6.16b, v6.16b, v16.16b\n" + "smax v7.16b, v7.16b, v16.16b\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "smax v23.16b, v5.16b, v4.16b\n" @@ -128,10 +128,10 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "smax v18.16b, v22.16b, v18.16b\n" "smax v17.16b, v21.16b, v17.16b\n" "smax v16.16b, v20.16b, v16.16b\n" - "smax v9.16b, v9.16b, v19.16b\n" - "smax v8.16b, v8.16b, v18.16b\n" - "smax v7.16b, v7.16b, v17.16b\n" - "smax v6.16b, v6.16b, v16.16b\n" + "smax v6.16b, v6.16b, v19.16b\n" + "smax v9.16b, v9.16b, v18.16b\n" + "smax v8.16b, v8.16b, v17.16b\n" + "smax v7.16b, v7.16b, v16.16b\n" "4:" // 4-vectors of channels: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" @@ -142,134 +142,155 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "ldr q18, [x20, x28]\n" "ldr q17, [x20, x27]\n" "ldr q16, [x20, x26]\n" - "smax v9.16b, v9.16b, v19.16b\n" - "smax v8.16b, v8.16b, v18.16b\n" - "smax v7.16b, v7.16b, v17.16b\n" - "smax v6.16b, v6.16b, v16.16b\n" + "smax v6.16b, v6.16b, v19.16b\n" + "smax v9.16b, v9.16b, v18.16b\n" + "smax v8.16b, v8.16b, v17.16b\n" + "smax v7.16b, v7.16b, v16.16b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "sxtl v23.8h, v9.8b\n" - "sxtl2 v19.8h, v9.16b\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" + "sxtl v23.8h, v6.8b\n" + "sxtl2 v19.8h, v6.16b\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1r { v6.4s }, [x21]\n" + "ld1r { v5.4s }, [x20]\n" + "sxtl v22.8h, v9.8b\n" + "sxtl2 v18.8h, v9.16b\n" + "sxtl v21.8h, v8.8b\n" + "sxtl2 v17.8h, v8.16b\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" "ld1r { v4.4s }, [x21]\n" "ld1r { v3.4s }, [x20]\n" - "sxtl v22.8h, v8.8b\n" - "sxtl2 v18.8h, v8.16b\n" - "sxtl v21.8h, v7.8b\n" - "sxtl2 v20.8h, v7.16b\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "sxtl v20.8h, v7.8b\n" + "sxtl2 v16.8h, v7.16b\n" + "neg v6.4s, v6.4s\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" + "movi v2.4s, #0x7f\n" "sub %x[n_channels], %x[n_channels], #0x40\n" - "ld1r { v2.4s }, [x20]\n" - "sxtl v17.8h, v6.8b\n" - "sxtl2 v16.8h, v6.16b\n" + "ld1r { v1.4s }, [x20]\n" "cmp %x[n_channels], #0x40\n" - "sxtl v1.4s, v23.4h\n" - "sxtl2 v23.4s, v23.8h\n" - "sxtl v0.4s, v19.4h\n" - "sxtl2 v19.4s, v19.8h\n" - "sxtl v31.4s, v22.4h\n" - "sxtl2 v22.4s, v22.8h\n" - "sxtl v30.4s, v18.4h\n" - "sxtl2 v18.4s, v18.8h\n" - "sxtl v29.4s, v21.4h\n" - "sxtl2 v21.4s, v21.8h\n" - "sxtl v28.4s, v20.4h\n" - "sxtl2 v27.4s, v20.8h\n" - "sxtl v26.4s, v17.4h\n" - "sxtl2 v20.4s, v17.8h\n" - "sxtl v25.4s, v16.4h\n" - "sxtl2 v24.4s, v16.8h\n" - "srshl v1.4s, v1.4s, v4.4s\n" - "srshl v23.4s, v23.4s, v4.4s\n" - "srshl v0.4s, v0.4s, v4.4s\n" - "srshl v19.4s, v19.4s, v4.4s\n" - "srshl v31.4s, v31.4s, v4.4s\n" - "srshl v22.4s, v22.4s, v4.4s\n" - "srshl v30.4s, v30.4s, v4.4s\n" - "srshl v18.4s, v18.4s, v4.4s\n" - "srshl v29.4s, v29.4s, v4.4s\n" - "srshl v21.4s, v21.4s, v4.4s\n" - "srshl v28.4s, v28.4s, v4.4s\n" - "srshl v27.4s, v27.4s, v4.4s\n" - "srshl v26.4s, v26.4s, v4.4s\n" - "srshl v20.4s, v20.4s, v4.4s\n" - "srshl v25.4s, v25.4s, v4.4s\n" - "srshl v24.4s, v24.4s, v4.4s\n" - "sqrdmulh v1.4s, v1.4s, v3.4s\n" - "sqrdmulh v23.4s, v23.4s, v3.4s\n" - "sqrdmulh v0.4s, v0.4s, v3.4s\n" - "sqrdmulh v19.4s, v19.4s, v3.4s\n" - "sqrdmulh v31.4s, v31.4s, v3.4s\n" - "sqrdmulh v22.4s, v22.4s, v3.4s\n" - "sqrdmulh v30.4s, v30.4s, v3.4s\n" - "sqrdmulh v18.4s, v18.4s, v3.4s\n" - "sqrdmulh v29.4s, v29.4s, v3.4s\n" - "sqrdmulh v21.4s, v21.4s, v3.4s\n" - "sqrdmulh v28.4s, v28.4s, v3.4s\n" - "sqrdmulh v27.4s, v27.4s, v3.4s\n" - "sqrdmulh v26.4s, v26.4s, v3.4s\n" - "sqrdmulh v20.4s, v20.4s, v3.4s\n" - "sqrdmulh v25.4s, v25.4s, v3.4s\n" - "sqrdmulh v24.4s, v24.4s, v3.4s\n" - "movi v17.4s, #0x7f\n" - "srshl v1.4s, v1.4s, v2.4s\n" - "srshl v23.4s, v23.4s, v2.4s\n" - "srshl v0.4s, v0.4s, v2.4s\n" - "srshl v19.4s, v19.4s, v2.4s\n" - "srshl v31.4s, v31.4s, v2.4s\n" - "srshl v22.4s, v22.4s, v2.4s\n" - "srshl v30.4s, v30.4s, v2.4s\n" - "srshl v18.4s, v18.4s, v2.4s\n" - "srshl v29.4s, v29.4s, v2.4s\n" - "srshl v21.4s, v21.4s, v2.4s\n" - "srshl v28.4s, v28.4s, v2.4s\n" - "srshl v27.4s, v27.4s, v2.4s\n" - "srshl v26.4s, v26.4s, v2.4s\n" - "srshl v20.4s, v20.4s, v2.4s\n" - "srshl v25.4s, v25.4s, v2.4s\n" - "srshl v24.4s, v24.4s, v2.4s\n" - "not v16.16b, v17.16b\n" - "smax v1.4s, v1.4s, v16.4s\n" - "smax v23.4s, v23.4s, v16.4s\n" - "smax v0.4s, v0.4s, v16.4s\n" - "smax v19.4s, v19.4s, v16.4s\n" - "smax v31.4s, v31.4s, v16.4s\n" - "smax v22.4s, v22.4s, v16.4s\n" - "smax v30.4s, v30.4s, v16.4s\n" - "smax v18.4s, v18.4s, v16.4s\n" - "smax v29.4s, v29.4s, v16.4s\n" - "smax v21.4s, v21.4s, v16.4s\n" - "smax v28.4s, v28.4s, v16.4s\n" - "smax v27.4s, v27.4s, v16.4s\n" - "smax v26.4s, v26.4s, v16.4s\n" - "smax v20.4s, v20.4s, v16.4s\n" - "smax v25.4s, v25.4s, v16.4s\n" - "smax v24.4s, v24.4s, v16.4s\n" - "smin v1.4s, v1.4s, v17.4s\n" - "smin v23.4s, v23.4s, v17.4s\n" - "smin v0.4s, v0.4s, v17.4s\n" - "smin v19.4s, v19.4s, v17.4s\n" - "smin v31.4s, v31.4s, v17.4s\n" - "smin v22.4s, v22.4s, v17.4s\n" - "smin v30.4s, v30.4s, v17.4s\n" - "smin v18.4s, v18.4s, v17.4s\n" - "smin v29.4s, v29.4s, v17.4s\n" - "smin v21.4s, v21.4s, v17.4s\n" - "smin v28.4s, v28.4s, v17.4s\n" - "smin v27.4s, v27.4s, v17.4s\n" - "smin v26.4s, v26.4s, v17.4s\n" - "smin v20.4s, v20.4s, v17.4s\n" - "smin v25.4s, v25.4s, v17.4s\n" - "smin v24.4s, v24.4s, v17.4s\n" - "uzp1 v23.16b, v1.16b, v23.16b\n" - "uzp1 v19.16b, v0.16b, v19.16b\n" - "uzp1 v22.16b, v31.16b, v22.16b\n" - "uzp1 v18.16b, v30.16b, v18.16b\n" - "uzp1 v21.16b, v29.16b, v21.16b\n" - "uzp1 v17.16b, v28.16b, v27.16b\n" - "uzp1 v20.16b, v26.16b, v20.16b\n" - "uzp1 v16.16b, v25.16b, v24.16b\n" + "not v0.16b, v2.16b\n" + "saddw v31.4s, v6.4s, v23.4h\n" + "saddw2 v23.4s, v6.4s, v23.8h\n" + "saddw v30.4s, v6.4s, v19.4h\n" + "saddw2 v19.4s, v6.4s, v19.8h\n" + "saddw v29.4s, v6.4s, v22.4h\n" + "saddw2 v22.4s, v6.4s, v22.8h\n" + "saddw v28.4s, v6.4s, v18.4h\n" + "saddw2 v18.4s, v6.4s, v18.8h\n" + "saddw v27.4s, v6.4s, v21.4h\n" + "saddw2 v21.4s, v6.4s, v21.8h\n" + "saddw v26.4s, v6.4s, v17.4h\n" + "saddw2 v17.4s, v6.4s, v17.8h\n" + "saddw v25.4s, v6.4s, v20.4h\n" + "saddw2 v20.4s, v6.4s, v20.8h\n" + "saddw v24.4s, v6.4s, v16.4h\n" + "saddw2 v16.4s, v6.4s, v16.8h\n" + "srshl v31.4s, v31.4s, v5.4s\n" + "srshl v23.4s, v23.4s, v5.4s\n" + "srshl v30.4s, v30.4s, v5.4s\n" + "srshl v19.4s, v19.4s, v5.4s\n" + "srshl v29.4s, v29.4s, v5.4s\n" + "srshl v22.4s, v22.4s, v5.4s\n" + "srshl v28.4s, v28.4s, v5.4s\n" + "srshl v18.4s, v18.4s, v5.4s\n" + "srshl v27.4s, v27.4s, v5.4s\n" + "srshl v21.4s, v21.4s, v5.4s\n" + "srshl v26.4s, v26.4s, v5.4s\n" + "srshl v17.4s, v17.4s, v5.4s\n" + "srshl v25.4s, v25.4s, v5.4s\n" + "srshl v20.4s, v20.4s, v5.4s\n" + "srshl v24.4s, v24.4s, v5.4s\n" + "srshl v16.4s, v16.4s, v5.4s\n" + "sqrdmulh v31.4s, v31.4s, v4.4s\n" + "sqrdmulh v23.4s, v23.4s, v4.4s\n" + "sqrdmulh v30.4s, v30.4s, v4.4s\n" + "sqrdmulh v19.4s, v19.4s, v4.4s\n" + "sqrdmulh v29.4s, v29.4s, v4.4s\n" + "sqrdmulh v22.4s, v22.4s, v4.4s\n" + "sqrdmulh v28.4s, v28.4s, v4.4s\n" + "sqrdmulh v18.4s, v18.4s, v4.4s\n" + "sqrdmulh v27.4s, v27.4s, v4.4s\n" + "sqrdmulh v21.4s, v21.4s, v4.4s\n" + "sqrdmulh v26.4s, v26.4s, v4.4s\n" + "sqrdmulh v17.4s, v17.4s, v4.4s\n" + "sqrdmulh v25.4s, v25.4s, v4.4s\n" + "sqrdmulh v20.4s, v20.4s, v4.4s\n" + "sqrdmulh v24.4s, v24.4s, v4.4s\n" + "sqrdmulh v16.4s, v16.4s, v4.4s\n" + "srshl v31.4s, v31.4s, v3.4s\n" + "srshl v23.4s, v23.4s, v3.4s\n" + "srshl v30.4s, v30.4s, v3.4s\n" + "srshl v19.4s, v19.4s, v3.4s\n" + "srshl v29.4s, v29.4s, v3.4s\n" + "srshl v22.4s, v22.4s, v3.4s\n" + "srshl v28.4s, v28.4s, v3.4s\n" + "srshl v18.4s, v18.4s, v3.4s\n" + "srshl v27.4s, v27.4s, v3.4s\n" + "srshl v21.4s, v21.4s, v3.4s\n" + "srshl v26.4s, v26.4s, v3.4s\n" + "srshl v17.4s, v17.4s, v3.4s\n" + "srshl v25.4s, v25.4s, v3.4s\n" + "srshl v20.4s, v20.4s, v3.4s\n" + "srshl v24.4s, v24.4s, v3.4s\n" + "srshl v16.4s, v16.4s, v3.4s\n" + "add v31.4s, v31.4s, v1.4s\n" + "add v23.4s, v23.4s, v1.4s\n" + "add v30.4s, v30.4s, v1.4s\n" + "add v19.4s, v19.4s, v1.4s\n" + "add v29.4s, v29.4s, v1.4s\n" + "add v22.4s, v22.4s, v1.4s\n" + "add v28.4s, v28.4s, v1.4s\n" + "add v18.4s, v18.4s, v1.4s\n" + "add v27.4s, v27.4s, v1.4s\n" + "add v21.4s, v21.4s, v1.4s\n" + "add v26.4s, v26.4s, v1.4s\n" + "add v17.4s, v17.4s, v1.4s\n" + "add v25.4s, v25.4s, v1.4s\n" + "add v20.4s, v20.4s, v1.4s\n" + "add v24.4s, v24.4s, v1.4s\n" + "add v16.4s, v16.4s, v1.4s\n" + "smax v31.4s, v31.4s, v0.4s\n" + "smax v23.4s, v23.4s, v0.4s\n" + "smax v30.4s, v30.4s, v0.4s\n" + "smax v19.4s, v19.4s, v0.4s\n" + "smax v29.4s, v29.4s, v0.4s\n" + "smax v22.4s, v22.4s, v0.4s\n" + "smax v28.4s, v28.4s, v0.4s\n" + "smax v18.4s, v18.4s, v0.4s\n" + "smax v27.4s, v27.4s, v0.4s\n" + "smax v21.4s, v21.4s, v0.4s\n" + "smax v26.4s, v26.4s, v0.4s\n" + "smax v17.4s, v17.4s, v0.4s\n" + "smax v25.4s, v25.4s, v0.4s\n" + "smax v20.4s, v20.4s, v0.4s\n" + "smax v24.4s, v24.4s, v0.4s\n" + "smax v16.4s, v16.4s, v0.4s\n" + "smin v31.4s, v31.4s, v2.4s\n" + "smin v23.4s, v23.4s, v2.4s\n" + "smin v30.4s, v30.4s, v2.4s\n" + "smin v19.4s, v19.4s, v2.4s\n" + "smin v29.4s, v29.4s, v2.4s\n" + "smin v22.4s, v22.4s, v2.4s\n" + "smin v28.4s, v28.4s, v2.4s\n" + "smin v18.4s, v18.4s, v2.4s\n" + "smin v27.4s, v27.4s, v2.4s\n" + "smin v21.4s, v21.4s, v2.4s\n" + "smin v26.4s, v26.4s, v2.4s\n" + "smin v17.4s, v17.4s, v2.4s\n" + "smin v25.4s, v25.4s, v2.4s\n" + "smin v20.4s, v20.4s, v2.4s\n" + "smin v24.4s, v24.4s, v2.4s\n" + "smin v16.4s, v16.4s, v2.4s\n" + "uzp1 v23.16b, v31.16b, v23.16b\n" + "uzp1 v19.16b, v30.16b, v19.16b\n" + "uzp1 v22.16b, v29.16b, v22.16b\n" + "uzp1 v18.16b, v28.16b, v18.16b\n" + "uzp1 v21.16b, v27.16b, v21.16b\n" + "uzp1 v17.16b, v26.16b, v17.16b\n" + "uzp1 v20.16b, v25.16b, v20.16b\n" + "uzp1 v16.16b, v24.16b, v16.16b\n" "uzp1 v19.16b, v23.16b, v19.16b\n" "uzp1 v18.16b, v22.16b, v18.16b\n" "uzp1 v17.16b, v21.16b, v17.16b\n" @@ -289,10 +310,10 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "blt 14f\n" "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" - "movi v9.16b, #0x80\n" + "movi v6.16b, #0x80\n" "mov x24, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "subs x25, x25, #0x1\n" "add x24, x24, #0x20\n" @@ -304,7 +325,7 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "9:" // Single vector of channels: Loop: 4 inputs loop "smax v17.16b, v5.16b, v4.16b\n" "smax v16.16b, v3.16b, v2.16b\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "subs x25, x25, #0x1\n" "add x24, x24, #0x20\n" @@ -313,13 +334,13 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "ldr q3, [x21, x9]\n" "ldr q2, [x20, x9]\n" "smax v16.16b, v17.16b, v16.16b\n" - "smax v9.16b, v9.16b, v16.16b\n" + "smax v6.16b, v6.16b, v16.16b\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail "smax v17.16b, v5.16b, v4.16b\n" "smax v16.16b, v3.16b, v2.16b\n" "smax v16.16b, v17.16b, v16.16b\n" - "smax v9.16b, v9.16b, v16.16b\n" + "smax v6.16b, v6.16b, v16.16b\n" "11:" // Single vector of channels: Loop: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 13f\n" @@ -327,47 +348,56 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "ldr x20, [x24], #0x8\n" "subs x21, x21, #0x1\n" "ldr q16, [x20, x9]\n" - "smax v9.16b, v9.16b, v16.16b\n" + "smax v6.16b, v6.16b, v16.16b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "sxtl v17.8h, v9.8b\n" - "sxtl2 v16.8h, v9.16b\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" + "sxtl v17.8h, v6.8b\n" + "sxtl2 v26.8h, v6.16b\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1r { v16.4s }, [x21]\n" + "ld1r { v25.4s }, [x20]\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" "ld1r { v24.4s }, [x21]\n" "ld1r { v23.4s }, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" "movi v22.4s, #0x7f\n" "ld1r { v21.4s }, [x20]\n" "sub %x[n_channels], %x[n_channels], #0x10\n" - "sxtl v20.4s, v17.4h\n" - "sxtl2 v17.4s, v17.8h\n" - "sxtl v19.4s, v16.4h\n" - "sxtl2 v18.4s, v16.8h\n" + "neg v16.4s, v16.4s\n" + "not v20.16b, v22.16b\n" "cmp %x[n_channels], #0x10\n" - "not v16.16b, v22.16b\n" - "srshl v20.4s, v20.4s, v24.4s\n" - "srshl v17.4s, v17.4s, v24.4s\n" - "srshl v19.4s, v19.4s, v24.4s\n" - "srshl v18.4s, v18.4s, v24.4s\n" - "sqrdmulh v20.4s, v20.4s, v23.4s\n" - "sqrdmulh v17.4s, v17.4s, v23.4s\n" - "sqrdmulh v19.4s, v19.4s, v23.4s\n" - "sqrdmulh v18.4s, v18.4s, v23.4s\n" - "srshl v20.4s, v20.4s, v21.4s\n" - "srshl v17.4s, v17.4s, v21.4s\n" - "srshl v19.4s, v19.4s, v21.4s\n" - "srshl v18.4s, v18.4s, v21.4s\n" - "smax v20.4s, v20.4s, v16.4s\n" - "smax v17.4s, v17.4s, v16.4s\n" - "smax v19.4s, v19.4s, v16.4s\n" - "smax v18.4s, v18.4s, v16.4s\n" - "smin v20.4s, v20.4s, v22.4s\n" - "smin v17.4s, v17.4s, v22.4s\n" + "saddw v19.4s, v16.4s, v17.4h\n" + "saddw2 v17.4s, v16.4s, v17.8h\n" + "saddw v18.4s, v16.4s, v26.4h\n" + "saddw2 v16.4s, v16.4s, v26.8h\n" + "srshl v19.4s, v19.4s, v25.4s\n" + "srshl v17.4s, v17.4s, v25.4s\n" + "srshl v18.4s, v18.4s, v25.4s\n" + "srshl v16.4s, v16.4s, v25.4s\n" + "sqrdmulh v19.4s, v19.4s, v24.4s\n" + "sqrdmulh v17.4s, v17.4s, v24.4s\n" + "sqrdmulh v18.4s, v18.4s, v24.4s\n" + "sqrdmulh v16.4s, v16.4s, v24.4s\n" + "srshl v19.4s, v19.4s, v23.4s\n" + "srshl v17.4s, v17.4s, v23.4s\n" + "srshl v18.4s, v18.4s, v23.4s\n" + "srshl v16.4s, v16.4s, v23.4s\n" + "add v19.4s, v19.4s, v21.4s\n" + "add v17.4s, v17.4s, v21.4s\n" + "add v18.4s, v18.4s, v21.4s\n" + "add v16.4s, v16.4s, v21.4s\n" + "smax v19.4s, v19.4s, v20.4s\n" + "smax v17.4s, v17.4s, v20.4s\n" + "smax v18.4s, v18.4s, v20.4s\n" + "smax v16.4s, v16.4s, v20.4s\n" "smin v19.4s, v19.4s, v22.4s\n" + "smin v17.4s, v17.4s, v22.4s\n" "smin v18.4s, v18.4s, v22.4s\n" - "uzp1 v17.16b, v20.16b, v17.16b\n" - "uzp1 v16.16b, v19.16b, v18.16b\n" + "smin v16.4s, v16.4s, v22.4s\n" + "uzp1 v17.16b, v19.16b, v17.16b\n" + "uzp1 v16.16b, v18.16b, v16.16b\n" "uzp1 v16.16b, v17.16b, v16.16b\n" "str q16, [%x[outptr], x9]\n" "add x9, x9, #0x10\n" @@ -376,17 +406,17 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "14:" // Oddments "lsr x25, %x[n_valid_cells], #0x2\n" "add %x[outptr], %x[outptr], x9\n" - "movi v9.16b, #0x80\n" + "movi v6.16b, #0x80\n" "mov x24, %x[inptrs]\n" "cbz x25, 24f\n" "15:" // Oddments: 4 inputs loop - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" - "movi v5.16b, #0x0\n" - "movi v4.16b, #0x0\n" - "movi v3.16b, #0x0\n" - "movi v2.16b, #0x0\n" + "movi v5.16b, #0\n" + "movi v4.16b, #0\n" + "movi v3.16b, #0\n" + "movi v2.16b, #0\n" "add x23, x23, x9\n" "add x22, x22, x9\n" "add x21, x21, x9\n" @@ -485,14 +515,14 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "smax v16.16b, v3.16b, v2.16b\n" "subs x25, x25, #0x1\n" "smax v16.16b, v17.16b, v16.16b\n" - "smax v9.16b, v9.16b, v16.16b\n" + "smax v6.16b, v6.16b, 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 x23, [x24], #0x8\n" - "movi v5.16b, #0x0\n" + "movi v5.16b, #0\n" "add x23, x23, x9\n" "tbz %x[n_channels], #3, 29f\n" "ldr d5, [x23], #0x8\n" @@ -540,45 +570,54 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "ldr b5, [x23], #0x1\n" "33:" // Oddments: Single input loop: Load: Bit 3: End "subs x21, x21, #0x1\n" - "smax v9.16b, v9.16b, v5.16b\n" + "smax v6.16b, v6.16b, v5.16b\n" "bgt 25b\n" "34:" // Oddments: Single input loop: End - "sxtl v17.8h, v9.8b\n" - "sxtl2 v16.8h, v9.16b\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" + "sxtl v17.8h, v6.8b\n" + "sxtl2 v26.8h, v6.16b\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1r { v16.4s }, [x21]\n" + "ld1r { v25.4s }, [x20]\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" "ld1r { v24.4s }, [x21]\n" "ld1r { v23.4s }, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" "movi v22.4s, #0x7f\n" "ld1r { v21.4s }, [x20]\n" - "sxtl v20.4s, v17.4h\n" - "sxtl2 v17.4s, v17.8h\n" - "sxtl v19.4s, v16.4h\n" - "sxtl2 v18.4s, v16.8h\n" - "not v16.16b, v22.16b\n" - "srshl v20.4s, v20.4s, v24.4s\n" - "srshl v17.4s, v17.4s, v24.4s\n" - "srshl v19.4s, v19.4s, v24.4s\n" - "srshl v18.4s, v18.4s, v24.4s\n" - "sqrdmulh v20.4s, v20.4s, v23.4s\n" - "sqrdmulh v17.4s, v17.4s, v23.4s\n" - "sqrdmulh v19.4s, v19.4s, v23.4s\n" - "sqrdmulh v18.4s, v18.4s, v23.4s\n" - "srshl v20.4s, v20.4s, v21.4s\n" - "srshl v17.4s, v17.4s, v21.4s\n" - "srshl v19.4s, v19.4s, v21.4s\n" - "srshl v18.4s, v18.4s, v21.4s\n" - "smax v20.4s, v20.4s, v16.4s\n" - "smax v17.4s, v17.4s, v16.4s\n" - "smax v19.4s, v19.4s, v16.4s\n" - "smax v18.4s, v18.4s, v16.4s\n" - "smin v20.4s, v20.4s, v22.4s\n" - "smin v17.4s, v17.4s, v22.4s\n" + "neg v16.4s, v16.4s\n" + "not v20.16b, v22.16b\n" + "saddw v19.4s, v16.4s, v17.4h\n" + "saddw2 v17.4s, v16.4s, v17.8h\n" + "saddw v18.4s, v16.4s, v26.4h\n" + "saddw2 v16.4s, v16.4s, v26.8h\n" + "srshl v19.4s, v19.4s, v25.4s\n" + "srshl v17.4s, v17.4s, v25.4s\n" + "srshl v18.4s, v18.4s, v25.4s\n" + "srshl v16.4s, v16.4s, v25.4s\n" + "sqrdmulh v19.4s, v19.4s, v24.4s\n" + "sqrdmulh v17.4s, v17.4s, v24.4s\n" + "sqrdmulh v18.4s, v18.4s, v24.4s\n" + "sqrdmulh v16.4s, v16.4s, v24.4s\n" + "srshl v19.4s, v19.4s, v23.4s\n" + "srshl v17.4s, v17.4s, v23.4s\n" + "srshl v18.4s, v18.4s, v23.4s\n" + "srshl v16.4s, v16.4s, v23.4s\n" + "add v19.4s, v19.4s, v21.4s\n" + "add v17.4s, v17.4s, v21.4s\n" + "add v18.4s, v18.4s, v21.4s\n" + "add v16.4s, v16.4s, v21.4s\n" + "smax v19.4s, v19.4s, v20.4s\n" + "smax v17.4s, v17.4s, v20.4s\n" + "smax v18.4s, v18.4s, v20.4s\n" + "smax v16.4s, v16.4s, v20.4s\n" "smin v19.4s, v19.4s, v22.4s\n" + "smin v17.4s, v17.4s, v22.4s\n" "smin v18.4s, v18.4s, v22.4s\n" - "uzp1 v17.16b, v20.16b, v17.16b\n" - "uzp1 v16.16b, v19.16b, v18.16b\n" + "smin v16.4s, v16.4s, v22.4s\n" + "uzp1 v17.16b, v19.16b, v17.16b\n" + "uzp1 v16.16b, v18.16b, v16.16b\n" "uzp1 v16.16b, v17.16b, v16.16b\n" "tbz %x[n_channels], #3, 38f\n" "st1 { v16.d }[0], [%x[outptr]], #0x8\n" @@ -627,7 +666,7 @@ void a64_s8q_nhwc_max_generic_depthfirst_impl( "42:" // Oddments: Store: Bit 3: End "43:" // End : [n_channels] "+&r" (n_channels), [outptr] "+&r" (outptr) - : [inptrs] "r" (inptrs), [n_valid_cells] "r" (n_valid_cells), [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) + : [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", "v9", "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" ); } diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp index 6895fd2011..1c91baf41b 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2024 Arm Limited. + * Copyright (c) 2022-2024, 2026 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,7 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( { __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "mov x9, #0x0\n" + "mov x9, #0\n" "cntb x28\n" "cntb x27, ALL, MUL #2\n" "cntb x26, ALL, MUL #3\n" @@ -52,16 +52,16 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "whilelt p2.b, x27, %x[n_channels]\n" "whilelt p1.b, x26, %x[n_channels]\n" "ptrue p0.b\n" - "b.none 7f\n" + "b.eq 7f\n" "1:" // 4-vectors of channels "lsr x25, %x[n_valid_cells], #0x2\n" "mov z5.b, #0x80\n" - "mov z3.b, #0x80\n" + "mov z4.b, #0x80\n" "mov x24, %x[inptrs]\n" + "mov z3.b, #0x80\n" "mov z2.b, #0x80\n" - "mov z4.b, #0x80\n" "cbz x25, 4f\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "subs x25, x25, #0x1\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" @@ -85,7 +85,7 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z1\n smax z19.b, p0/M, z19.b, z0.b\n" "smax z23.b, p0/M, z23.b, z31.b\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "subs x25, x25, #0x1\n" "movprfx z18, z30\n smax z18.b, p0/M, z18.b, z29.b\n" "smax z22.b, p0/M, z22.b, z28.b\n" @@ -106,11 +106,11 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "ld1b { z30.b }, p3/Z, [x23, x28]\n" "smax z5.b, p0/M, z5.b, z19.b\n" "ld1b { z29.b }, p3/Z, [x22, x28]\n" - "smax z3.b, p0/M, z3.b, z18.b\n" + "smax z4.b, p0/M, z4.b, z18.b\n" "ld1b { z22.b }, p3/Z, [x21, x28]\n" - "smax z2.b, p0/M, z2.b, z17.b\n" + "smax z3.b, p0/M, z3.b, z17.b\n" "ld1b { z28.b }, p3/Z, [x20, x28]\n" - "smax z4.b, p0/M, z4.b, z16.b\n" + "smax z2.b, p0/M, z2.b, z16.b\n" "ld1b { z17.b }, p2/Z, [x23, x27]\n" "ld1b { z27.b }, p2/Z, [x22, x27]\n" "ld1b { z21.b }, p2/Z, [x21, x27]\n" @@ -134,9 +134,9 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "smax z17.b, p0/M, z17.b, z21.b\n" "smax z16.b, p0/M, z16.b, z20.b\n" "smax z5.b, p0/M, z5.b, z19.b\n" - "smax z3.b, p0/M, z3.b, z18.b\n" - "smax z2.b, p0/M, z2.b, z17.b\n" - "smax z4.b, p0/M, z4.b, z16.b\n" + "smax z4.b, p0/M, z4.b, z18.b\n" + "smax z3.b, p0/M, z3.b, z17.b\n" + "smax z2.b, p0/M, z2.b, z16.b\n" "4:" // 4-vectors of channels: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" @@ -148,129 +148,150 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "ld1b { z17.b }, p2/Z, [x20, x27]\n" "ld1b { z16.b }, p1/Z, [x20, x26]\n" "smax z5.b, p0/M, z5.b, z19.b\n" - "smax z3.b, p0/M, z3.b, z18.b\n" - "smax z2.b, p0/M, z2.b, z17.b\n" - "smax z4.b, p0/M, z4.b, z16.b\n" + "smax z4.b, p0/M, z4.b, z18.b\n" + "smax z3.b, p0/M, z3.b, z17.b\n" + "smax z2.b, p0/M, z2.b, z16.b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - ".inst 0x4508a0b3 // sshllb z19.h, z5.b, #0x0\n" - ".inst 0x4508a4b8 // sshllt z24.h, z5.b, #0x0\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - ".inst 0x4508a076 // sshllb z22.h, z3.b, #0x0\n" - ".inst 0x4508a472 // sshllt z18.h, z3.b, #0x0\n" + "add x20, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a0b3 // sshllb z19.h, z5.b, #0\n" + ".inst 0x4508a4b6 // sshllt z22.h, z5.b, #0\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z5.s }, p0/Z, [x20]\n" + ".inst 0x4508a095 // sshllb z21.h, z4.b, #0\n" + ".inst 0x4508a492 // sshllt z18.h, z4.b, #0\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + ".inst 0x4508a074 // sshllb z20.h, z3.b, #0\n" + ".inst 0x4508a471 // sshllt z17.h, z3.b, #0\n" + "ld1rw { z4.s }, p0/Z, [x21]\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + ".inst 0x4508a058 // sshllb z24.h, z2.b, #0\n" + ".inst 0x4508a457 // sshllt z23.h, z2.b, #0\n" "ld1rw { z3.s }, p0/Z, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - ".inst 0x4508a055 // sshllb z21.h, z2.b, #0x0\n" - ".inst 0x4508a454 // sshllt z20.h, z2.b, #0x0\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" + "neg z5.s, p0/M, z5.s\n" "ld1rw { z2.s }, p0/Z, [x21]\n" - ".inst 0x4508a097 // sshllb z23.h, z4.b, #0x0\n" - ".inst 0x4508a491 // sshllt z17.h, z4.b, #0x0\n" + "mov z1.s, #0x7f\n" "ld1rw { z16.s }, p0/Z, [x20]\n" - ".inst 0x4510a261 // sshllb z1.s, z19.h, #0x0\n" - ".inst 0x4510a673 // sshllt z19.s, z19.h, #0x0\n" - ".inst 0x4510a300 // sshllb z0.s, z24.h, #0x0\n" - ".inst 0x4510a71f // sshllt z31.s, z24.h, #0x0\n" - ".inst 0x4510a2de // sshllb z30.s, z22.h, #0x0\n" - ".inst 0x4510a6d6 // sshllt z22.s, z22.h, #0x0\n" - ".inst 0x4510a25d // sshllb z29.s, z18.h, #0x0\n" - ".inst 0x4510a652 // sshllt z18.s, z18.h, #0x0\n" - ".inst 0x4510a2bc // sshllb z28.s, z21.h, #0x0\n" - ".inst 0x4510a6b5 // sshllt z21.s, z21.h, #0x0\n" - ".inst 0x4510a29b // sshllb z27.s, z20.h, #0x0\n" - ".inst 0x4510a694 // sshllt z20.s, z20.h, #0x0\n" - ".inst 0x4510a2fa // sshllb z26.s, z23.h, #0x0\n" - ".inst 0x4510a6f9 // sshllt z25.s, z23.h, #0x0\n" - ".inst 0x4510a238 // sshllb z24.s, z17.h, #0x0\n" - ".inst 0x4510a637 // sshllt z23.s, z17.h, #0x0\n" - ".inst 0x44828061 // srshl z1.s, p0/M, z1.s, z3.s\n" - ".inst 0x44828073 // srshl z19.s, p0/M, z19.s, z3.s\n" - ".inst 0x44828060 // srshl z0.s, p0/M, z0.s, z3.s\n" - ".inst 0x4482807f // srshl z31.s, p0/M, z31.s, z3.s\n" - ".inst 0x4482807e // srshl z30.s, p0/M, z30.s, z3.s\n" - ".inst 0x44828076 // srshl z22.s, p0/M, z22.s, z3.s\n" - ".inst 0x4482807d // srshl z29.s, p0/M, z29.s, z3.s\n" - ".inst 0x44828072 // srshl z18.s, p0/M, z18.s, z3.s\n" - ".inst 0x4482807c // srshl z28.s, p0/M, z28.s, z3.s\n" - ".inst 0x44828075 // srshl z21.s, p0/M, z21.s, z3.s\n" - ".inst 0x4482807b // srshl z27.s, p0/M, z27.s, z3.s\n" - ".inst 0x44828074 // srshl z20.s, p0/M, z20.s, z3.s\n" - ".inst 0x4482807a // srshl z26.s, p0/M, z26.s, z3.s\n" - ".inst 0x44828079 // srshl z25.s, p0/M, z25.s, z3.s\n" - ".inst 0x44828078 // srshl z24.s, p0/M, z24.s, z3.s\n" - ".inst 0x44828077 // srshl z23.s, p0/M, z23.s, z3.s\n" - ".inst 0x04a27421 // sqrdmulh z1.s, z1.s, z2.s\n" - ".inst 0x04a27673 // sqrdmulh z19.s, z19.s, z2.s\n" - ".inst 0x04a27400 // sqrdmulh z0.s, z0.s, z2.s\n" - ".inst 0x04a277ff // sqrdmulh z31.s, z31.s, z2.s\n" - ".inst 0x04a277de // sqrdmulh z30.s, z30.s, z2.s\n" - ".inst 0x04a276d6 // sqrdmulh z22.s, z22.s, z2.s\n" - ".inst 0x04a277bd // sqrdmulh z29.s, z29.s, z2.s\n" - ".inst 0x04a27652 // sqrdmulh z18.s, z18.s, z2.s\n" - ".inst 0x04a2779c // sqrdmulh z28.s, z28.s, z2.s\n" - ".inst 0x04a276b5 // sqrdmulh z21.s, z21.s, z2.s\n" - ".inst 0x04a2777b // sqrdmulh z27.s, z27.s, z2.s\n" - ".inst 0x04a27694 // sqrdmulh z20.s, z20.s, z2.s\n" - ".inst 0x04a2775a // sqrdmulh z26.s, z26.s, z2.s\n" - ".inst 0x04a27739 // sqrdmulh z25.s, z25.s, z2.s\n" - ".inst 0x04a27718 // sqrdmulh z24.s, z24.s, z2.s\n" - ".inst 0x04a276f7 // sqrdmulh z23.s, z23.s, z2.s\n" - "mov z17.s, #0x7f\n" - ".inst 0x44828201 // srshl z1.s, p0/M, z1.s, z16.s\n" - ".inst 0x44828213 // srshl z19.s, p0/M, z19.s, z16.s\n" - ".inst 0x44828200 // srshl z0.s, p0/M, z0.s, z16.s\n" - ".inst 0x4482821f // srshl z31.s, p0/M, z31.s, z16.s\n" - ".inst 0x4482821e // srshl z30.s, p0/M, z30.s, z16.s\n" - ".inst 0x44828216 // srshl z22.s, p0/M, z22.s, z16.s\n" - ".inst 0x4482821d // srshl z29.s, p0/M, z29.s, z16.s\n" - ".inst 0x44828212 // srshl z18.s, p0/M, z18.s, z16.s\n" - ".inst 0x4482821c // srshl z28.s, p0/M, z28.s, z16.s\n" - ".inst 0x44828215 // srshl z21.s, p0/M, z21.s, z16.s\n" - ".inst 0x4482821b // srshl z27.s, p0/M, z27.s, z16.s\n" - ".inst 0x44828214 // srshl z20.s, p0/M, z20.s, z16.s\n" - ".inst 0x4482821a // srshl z26.s, p0/M, z26.s, z16.s\n" - ".inst 0x44828219 // srshl z25.s, p0/M, z25.s, z16.s\n" - ".inst 0x44828218 // srshl z24.s, p0/M, z24.s, z16.s\n" - ".inst 0x44828217 // srshl z23.s, p0/M, z23.s, z16.s\n" - "not z16.s, p0/M, z17.s\n" - "smax z1.s, p0/M, z1.s, z16.s\n" - "smax z19.s, p0/M, z19.s, z16.s\n" + ".inst 0x459340a0 // saddwb z0.s, z5.s, z19.h\n" + ".inst 0x459344b3 // saddwt z19.s, z5.s, z19.h\n" + ".inst 0x459640bf // saddwb z31.s, z5.s, z22.h\n" + ".inst 0x459644be // saddwt z30.s, z5.s, z22.h\n" + ".inst 0x459540bd // saddwb z29.s, z5.s, z21.h\n" + ".inst 0x459544b6 // saddwt z22.s, z5.s, z21.h\n" + ".inst 0x459240bc // saddwb z28.s, z5.s, z18.h\n" + ".inst 0x459244b2 // saddwt z18.s, z5.s, z18.h\n" + ".inst 0x459440bb // saddwb z27.s, z5.s, z20.h\n" + ".inst 0x459444b5 // saddwt z21.s, z5.s, z20.h\n" + ".inst 0x459140b4 // saddwb z20.s, z5.s, z17.h\n" + ".inst 0x459144b1 // saddwt z17.s, z5.s, z17.h\n" + ".inst 0x459840ba // saddwb z26.s, z5.s, z24.h\n" + ".inst 0x459844b9 // saddwt z25.s, z5.s, z24.h\n" + ".inst 0x459740b8 // saddwb z24.s, z5.s, z23.h\n" + ".inst 0x459744b7 // saddwt z23.s, z5.s, z23.h\n" + ".inst 0x44828080 // srshl z0.s, p0/M, z0.s, z4.s\n" + ".inst 0x44828093 // srshl z19.s, p0/M, z19.s, z4.s\n" + ".inst 0x4482809f // srshl z31.s, p0/M, z31.s, z4.s\n" + ".inst 0x4482809e // srshl z30.s, p0/M, z30.s, z4.s\n" + ".inst 0x4482809d // srshl z29.s, p0/M, z29.s, z4.s\n" + ".inst 0x44828096 // srshl z22.s, p0/M, z22.s, z4.s\n" + ".inst 0x4482809c // srshl z28.s, p0/M, z28.s, z4.s\n" + ".inst 0x44828092 // srshl z18.s, p0/M, z18.s, z4.s\n" + ".inst 0x4482809b // srshl z27.s, p0/M, z27.s, z4.s\n" + ".inst 0x44828095 // srshl z21.s, p0/M, z21.s, z4.s\n" + ".inst 0x44828094 // srshl z20.s, p0/M, z20.s, z4.s\n" + ".inst 0x44828091 // srshl z17.s, p0/M, z17.s, z4.s\n" + ".inst 0x4482809a // srshl z26.s, p0/M, z26.s, z4.s\n" + ".inst 0x44828099 // srshl z25.s, p0/M, z25.s, z4.s\n" + ".inst 0x44828098 // srshl z24.s, p0/M, z24.s, z4.s\n" + ".inst 0x44828097 // srshl z23.s, p0/M, z23.s, z4.s\n" + ".inst 0x04a37400 // sqrdmulh z0.s, z0.s, z3.s\n" + ".inst 0x04a37673 // sqrdmulh z19.s, z19.s, z3.s\n" + ".inst 0x04a377ff // sqrdmulh z31.s, z31.s, z3.s\n" + ".inst 0x04a377de // sqrdmulh z30.s, z30.s, z3.s\n" + ".inst 0x04a377bd // sqrdmulh z29.s, z29.s, z3.s\n" + ".inst 0x04a376d6 // sqrdmulh z22.s, z22.s, z3.s\n" + ".inst 0x04a3779c // sqrdmulh z28.s, z28.s, z3.s\n" + ".inst 0x04a37652 // sqrdmulh z18.s, z18.s, z3.s\n" + ".inst 0x04a3777b // sqrdmulh z27.s, z27.s, z3.s\n" + ".inst 0x04a376b5 // sqrdmulh z21.s, z21.s, z3.s\n" + ".inst 0x04a37694 // sqrdmulh z20.s, z20.s, z3.s\n" + ".inst 0x04a37631 // sqrdmulh z17.s, z17.s, z3.s\n" + ".inst 0x04a3775a // sqrdmulh z26.s, z26.s, z3.s\n" + ".inst 0x04a37739 // sqrdmulh z25.s, z25.s, z3.s\n" + ".inst 0x04a37718 // sqrdmulh z24.s, z24.s, z3.s\n" + ".inst 0x04a376f7 // sqrdmulh z23.s, z23.s, z3.s\n" + ".inst 0x44828040 // srshl z0.s, p0/M, z0.s, z2.s\n" + ".inst 0x44828053 // srshl z19.s, p0/M, z19.s, z2.s\n" + ".inst 0x4482805f // srshl z31.s, p0/M, z31.s, z2.s\n" + ".inst 0x4482805e // srshl z30.s, p0/M, z30.s, z2.s\n" + ".inst 0x4482805d // srshl z29.s, p0/M, z29.s, z2.s\n" + ".inst 0x44828056 // srshl z22.s, p0/M, z22.s, z2.s\n" + ".inst 0x4482805c // srshl z28.s, p0/M, z28.s, z2.s\n" + ".inst 0x44828052 // srshl z18.s, p0/M, z18.s, z2.s\n" + ".inst 0x4482805b // srshl z27.s, p0/M, z27.s, z2.s\n" + ".inst 0x44828055 // srshl z21.s, p0/M, z21.s, z2.s\n" + ".inst 0x44828054 // srshl z20.s, p0/M, z20.s, z2.s\n" + ".inst 0x44828051 // srshl z17.s, p0/M, z17.s, z2.s\n" + ".inst 0x4482805a // srshl z26.s, p0/M, z26.s, z2.s\n" + ".inst 0x44828059 // srshl z25.s, p0/M, z25.s, z2.s\n" + ".inst 0x44828058 // srshl z24.s, p0/M, z24.s, z2.s\n" + ".inst 0x44828057 // srshl z23.s, p0/M, z23.s, z2.s\n" + "add z0.s, z0.s, z16.s\n" + "add z19.s, z19.s, z16.s\n" + "add z31.s, z31.s, z16.s\n" + "add z30.s, z30.s, z16.s\n" + "add z29.s, z29.s, z16.s\n" + "add z22.s, z22.s, z16.s\n" + "add z28.s, z28.s, z16.s\n" + "add z18.s, z18.s, z16.s\n" + "add z27.s, z27.s, z16.s\n" + "add z21.s, z21.s, z16.s\n" + "add z20.s, z20.s, z16.s\n" + "add z17.s, z17.s, z16.s\n" + "add z26.s, z26.s, z16.s\n" + "add z25.s, z25.s, z16.s\n" + "add z24.s, z24.s, z16.s\n" + "add z23.s, z23.s, z16.s\n" + "not z16.s, p0/M, z1.s\n" "smax z0.s, p0/M, z0.s, z16.s\n" + "smax z19.s, p0/M, z19.s, z16.s\n" "smax z31.s, p0/M, z31.s, z16.s\n" "smax z30.s, p0/M, z30.s, z16.s\n" - "smax z22.s, p0/M, z22.s, z16.s\n" "smax z29.s, p0/M, z29.s, z16.s\n" - "smax z18.s, p0/M, z18.s, z16.s\n" + "smax z22.s, p0/M, z22.s, z16.s\n" "smax z28.s, p0/M, z28.s, z16.s\n" - "smax z21.s, p0/M, z21.s, z16.s\n" + "smax z18.s, p0/M, z18.s, z16.s\n" "smax z27.s, p0/M, z27.s, z16.s\n" + "smax z21.s, p0/M, z21.s, z16.s\n" "smax z20.s, p0/M, z20.s, z16.s\n" + "smax z17.s, p0/M, z17.s, z16.s\n" "smax z26.s, p0/M, z26.s, z16.s\n" "smax z25.s, p0/M, z25.s, z16.s\n" "smax z24.s, p0/M, z24.s, z16.s\n" "smax z23.s, p0/M, z23.s, z16.s\n" - "smin z1.s, p0/M, z1.s, z17.s\n" - "smin z19.s, p0/M, z19.s, z17.s\n" - "smin z0.s, p0/M, z0.s, z17.s\n" - "smin z31.s, p0/M, z31.s, z17.s\n" - "smin z30.s, p0/M, z30.s, z17.s\n" - "smin z22.s, p0/M, z22.s, z17.s\n" - "smin z29.s, p0/M, z29.s, z17.s\n" - "smin z18.s, p0/M, z18.s, z17.s\n" - "smin z28.s, p0/M, z28.s, z17.s\n" - "smin z21.s, p0/M, z21.s, z17.s\n" - "trn1 z19.h, z1.h, z19.h\n" - "smin z27.s, p0/M, z27.s, z17.s\n" - "smin z20.s, p0/M, z20.s, z17.s\n" - "trn1 z16.h, z0.h, z31.h\n" - "smin z26.s, p0/M, z26.s, z17.s\n" - "smin z25.s, p0/M, z25.s, z17.s\n" - "trn1 z22.h, z30.h, z22.h\n" - "smin z24.s, p0/M, z24.s, z17.s\n" - "smin z23.s, p0/M, z23.s, z17.s\n" - "trn1 z18.h, z29.h, z18.h\n" - "trn1 z21.h, z28.h, z21.h\n" - "trn1 z17.h, z27.h, z20.h\n" + "smin z0.s, p0/M, z0.s, z1.s\n" + "smin z19.s, p0/M, z19.s, z1.s\n" + "smin z31.s, p0/M, z31.s, z1.s\n" + "smin z30.s, p0/M, z30.s, z1.s\n" + "smin z29.s, p0/M, z29.s, z1.s\n" + "smin z22.s, p0/M, z22.s, z1.s\n" + "smin z28.s, p0/M, z28.s, z1.s\n" + "smin z18.s, p0/M, z18.s, z1.s\n" + "smin z27.s, p0/M, z27.s, z1.s\n" + "smin z21.s, p0/M, z21.s, z1.s\n" + "trn1 z19.h, z0.h, z19.h\n" + "smin z20.s, p0/M, z20.s, z1.s\n" + "smin z17.s, p0/M, z17.s, z1.s\n" + "trn1 z16.h, z31.h, z30.h\n" + "smin z26.s, p0/M, z26.s, z1.s\n" + "smin z25.s, p0/M, z25.s, z1.s\n" + "trn1 z22.h, z29.h, z22.h\n" + "smin z24.s, p0/M, z24.s, z1.s\n" + "smin z23.s, p0/M, z23.s, z1.s\n" + "trn1 z18.h, z28.h, z18.h\n" + "trn1 z21.h, z27.h, z21.h\n" + "trn1 z17.h, z20.h, z17.h\n" "trn1 z20.b, z19.b, z16.b\n" "trn1 z19.h, z26.h, z25.h\n" "trn1 z16.h, z24.h, z23.h\n" @@ -286,16 +307,16 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "st1b { z16.b }, p1, [%x[outptr], x26]\n" "incb x26, ALL, MUL #4\n" "whilelt p1.b, x26, %x[n_channels]\n" - "b.any 1b\n" + "b.ne 1b\n" "7:" // Single vector of channels "whilelt p4.b, x9, %x[n_channels]\n" - "b.none 14f\n" + "b.eq 14f\n" "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" "mov z5.b, #0x80\n" "mov x24, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "subs x25, x25, #0x1\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" @@ -307,7 +328,7 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "9:" // Single vector of channels: Loop: 4 inputs loop "movprfx z16, z1\n smax z16.b, p0/M, z16.b, z0.b\n" "movprfx z17, z23\n smax z17.b, p0/M, z17.b, z31.b\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "subs x25, x25, #0x1\n" "ldp x21, x20, [x24, #0x10]\n" "add x24, x24, #0x20\n" @@ -333,51 +354,60 @@ void sme_s8q_nhwc_max_generic_depthfirst_impl( "smax z5.b, p0/M, z5.b, z16.b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - ".inst 0x4508a0b1 // sshllb z17.h, z5.b, #0x0\n" - ".inst 0x4508a4b0 // sshllt z16.h, z5.b, #0x0\n" + "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a0b1 // sshllb z17.h, z5.b, #0\n" + ".inst 0x4508a4ba // sshllt z26.h, z5.b, #0\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "ld1rw { z16.s }, p0/Z, [x21]\n" + "add x22, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "mov z25.s, #0x7f\n" "ld1rw { z24.s }, p0/Z, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - "mov z23.s, #0x7f\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" + "ld1rw { z23.s }, p0/Z, [x22]\n" + "neg z16.s, p0/M, z16.s\n" "ld1rw { z22.s }, p0/Z, [x21]\n" "ld1rw { z21.s }, p0/Z, [x20]\n" - ".inst 0x4510a234 // sshllb z20.s, z17.h, #0x0\n" - ".inst 0x4510a631 // sshllt z17.s, z17.h, #0x0\n" - ".inst 0x4510a213 // sshllb z19.s, z16.h, #0x0\n" - ".inst 0x4510a612 // sshllt z18.s, z16.h, #0x0\n" - "not z16.s, p0/M, z23.s\n" - ".inst 0x44828314 // srshl z20.s, p0/M, z20.s, z24.s\n" - ".inst 0x44828311 // srshl z17.s, p0/M, z17.s, z24.s\n" + "not z20.s, p0/M, z25.s\n" + ".inst 0x45914213 // saddwb z19.s, z16.s, z17.h\n" + ".inst 0x45914611 // saddwt z17.s, z16.s, z17.h\n" + ".inst 0x459a4212 // saddwb z18.s, z16.s, z26.h\n" + ".inst 0x459a4610 // saddwt z16.s, z16.s, z26.h\n" ".inst 0x44828313 // srshl z19.s, p0/M, z19.s, z24.s\n" + ".inst 0x44828311 // srshl z17.s, p0/M, z17.s, z24.s\n" ".inst 0x44828312 // srshl z18.s, p0/M, z18.s, z24.s\n" - ".inst 0x04b67694 // sqrdmulh z20.s, z20.s, z22.s\n" - ".inst 0x04b67631 // sqrdmulh z17.s, z17.s, z22.s\n" - ".inst 0x04b67673 // sqrdmulh z19.s, z19.s, z22.s\n" - ".inst 0x04b67652 // sqrdmulh z18.s, z18.s, z22.s\n" - ".inst 0x448282b4 // srshl z20.s, p0/M, z20.s, z21.s\n" - ".inst 0x448282b1 // srshl z17.s, p0/M, z17.s, z21.s\n" - ".inst 0x448282b3 // srshl z19.s, p0/M, z19.s, z21.s\n" - ".inst 0x448282b2 // srshl z18.s, p0/M, z18.s, z21.s\n" - "smax z20.s, p0/M, z20.s, z16.s\n" - "smax z17.s, p0/M, z17.s, z16.s\n" - "smax z19.s, p0/M, z19.s, z16.s\n" - "smax z18.s, p0/M, z18.s, z16.s\n" - "smin z20.s, p0/M, z20.s, z23.s\n" - "smin z17.s, p0/M, z17.s, z23.s\n" - "smin z19.s, p0/M, z19.s, z23.s\n" - "smin z18.s, p0/M, z18.s, z23.s\n" - "trn1 z17.h, z20.h, z17.h\n" - "trn1 z16.h, z19.h, z18.h\n" + ".inst 0x44828310 // srshl z16.s, p0/M, z16.s, z24.s\n" + ".inst 0x04b77673 // sqrdmulh z19.s, z19.s, z23.s\n" + ".inst 0x04b77631 // sqrdmulh z17.s, z17.s, z23.s\n" + ".inst 0x04b77652 // sqrdmulh z18.s, z18.s, z23.s\n" + ".inst 0x04b77610 // sqrdmulh z16.s, z16.s, z23.s\n" + ".inst 0x448282d3 // srshl z19.s, p0/M, z19.s, z22.s\n" + ".inst 0x448282d1 // srshl z17.s, p0/M, z17.s, z22.s\n" + ".inst 0x448282d2 // srshl z18.s, p0/M, z18.s, z22.s\n" + ".inst 0x448282d0 // srshl z16.s, p0/M, z16.s, z22.s\n" + "add z19.s, z19.s, z21.s\n" + "add z17.s, z17.s, z21.s\n" + "add z18.s, z18.s, z21.s\n" + "add z16.s, z16.s, z21.s\n" + "smax z19.s, p0/M, z19.s, z20.s\n" + "smax z17.s, p0/M, z17.s, z20.s\n" + "smax z18.s, p0/M, z18.s, z20.s\n" + "smax z16.s, p0/M, z16.s, z20.s\n" + "smin z19.s, p0/M, z19.s, z25.s\n" + "smin z17.s, p0/M, z17.s, z25.s\n" + "smin z18.s, p0/M, z18.s, z25.s\n" + "smin z16.s, p0/M, z16.s, z25.s\n" + "trn1 z17.h, z19.h, z17.h\n" + "trn1 z16.h, z18.h, z16.h\n" "trn1 z16.b, z17.b, z16.b\n" "st1b { z16.b }, p4, [%x[outptr], x9]\n" "incb x9\n" "whilelt p4.b, x9, %x[n_channels]\n" - "b.any 8b\n" + "b.ne 8b\n" "14:" // End ".inst 0xd503467f // SMSTOP\n" : - : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [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)), [outptr] "r" (outptr), [quant_params] "r" (&qp) + : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [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)), [outptr] "r" (outptr), [quant_params] "r" (&qp) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp index 5033aa9d73..0b3f4b88ec 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_s8q_nhwc_max_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2024 Arm Limited. + * Copyright (c) 2021-2024, 2026 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,7 +42,7 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( ) { __asm__ __volatile__( - "mov x9, #0x0\n" + "mov x9, #0\n" "cntb x28\n" "cntb x27, ALL, MUL #2\n" "cntb x26, ALL, MUL #3\n" @@ -51,16 +51,16 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "whilelt p2.b, x28, %x[n_channels]\n" "whilelt p1.b, x27, %x[n_channels]\n" "whilelt p0.b, x26, %x[n_channels]\n" - "b.none 7f\n" + "b.eq 7f\n" "1:" // 4-vectors of channels "lsr x25, %x[n_valid_cells], #0x2\n" "mov z6.b, #0x80\n" - "mov z3.b, #0x80\n" - "mov x24, %x[inptrs]\n" "mov z5.b, #0x80\n" + "mov x24, %x[inptrs]\n" "mov z4.b, #0x80\n" + "mov z3.b, #0x80\n" "cbz x25, 4f\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "subs x25, x25, #0x1\n" "add x24, x24, #0x20\n" @@ -84,7 +84,7 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z2\n smax z19.b, p4/M, z19.b, z1.b\n" "smax z23.b, p4/M, z23.b, z0.b\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "movprfx z18, z31\n smax z18.b, p4/M, z18.b, z30.b\n" "smax z22.b, p4/M, z22.b, z29.b\n" @@ -107,13 +107,13 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "ld1b { z22.b }, p2/Z, [x21, x28]\n" "ld1b { z29.b }, p2/Z, [x20, x28]\n" "smax z6.b, p4/M, z6.b, z19.b\n" - "smax z3.b, p4/M, z3.b, z18.b\n" + "smax z5.b, p4/M, z5.b, z18.b\n" "ld1b { z28.b }, p1/Z, [x23, x27]\n" "ld1b { z27.b }, p1/Z, [x22, x27]\n" - "smax z5.b, p4/M, z5.b, z17.b\n" + "smax z4.b, p4/M, z4.b, z17.b\n" "ld1b { z21.b }, p1/Z, [x21, x27]\n" "ld1b { z26.b }, p1/Z, [x20, x27]\n" - "smax z4.b, p4/M, z4.b, z16.b\n" + "smax z3.b, p4/M, z3.b, z16.b\n" "ld1b { z16.b }, p0/Z, [x23, x26]\n" "ld1b { z25.b }, p0/Z, [x22, x26]\n" "ld1b { z20.b }, p0/Z, [x21, x26]\n" @@ -133,9 +133,9 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smax z17.b, p4/M, z17.b, z21.b\n" "smax z16.b, p4/M, z16.b, z20.b\n" "smax z6.b, p4/M, z6.b, z19.b\n" - "smax z3.b, p4/M, z3.b, z18.b\n" - "smax z5.b, p4/M, z5.b, z17.b\n" - "smax z4.b, p4/M, z4.b, z16.b\n" + "smax z5.b, p4/M, z5.b, z18.b\n" + "smax z4.b, p4/M, z4.b, z17.b\n" + "smax z3.b, p4/M, z3.b, z16.b\n" "4:" // 4-vectors of channels: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" @@ -147,128 +147,149 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "ld1b { z17.b }, p1/Z, [x20, x27]\n" "ld1b { z16.b }, p0/Z, [x20, x26]\n" "smax z6.b, p4/M, z6.b, z19.b\n" - "smax z3.b, p4/M, z3.b, z18.b\n" - "smax z5.b, p4/M, z5.b, z17.b\n" - "smax z4.b, p4/M, z4.b, z16.b\n" + "smax z5.b, p4/M, z5.b, z18.b\n" + "smax z4.b, p4/M, z4.b, z17.b\n" + "smax z3.b, p4/M, z3.b, z16.b\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - ".inst 0x4508a0d3 // sshllb z19.h, z6.b, #0x0\n" - ".inst 0x4508a4d1 // sshllt z17.h, z6.b, #0x0\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" - ".inst 0x4508a072 // sshllb z18.h, z3.b, #0x0\n" - ".inst 0x4508a478 // sshllt z24.h, z3.b, #0x0\n" - "ld1rw { z3.s }, p4/Z, [x21]\n" - "ld1rw { z2.s }, p4/Z, [x20]\n" - ".inst 0x4508a0b5 // sshllb z21.h, z5.b, #0x0\n" - ".inst 0x4508a4b7 // sshllt z23.h, z5.b, #0x0\n" + "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a0d3 // sshllb z19.h, z6.b, #0\n" + ".inst 0x4508a4d1 // sshllt z17.h, z6.b, #0\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z6.s }, p4/Z, [x21]\n" + ".inst 0x4508a0b2 // sshllb z18.h, z5.b, #0\n" + ".inst 0x4508a4b0 // sshllt z16.h, z5.b, #0\n" + "ld1rw { z5.s }, p4/Z, [x20]\n" + ".inst 0x4508a094 // sshllb z20.h, z4.b, #0\n" + ".inst 0x4508a498 // sshllt z24.h, z4.b, #0\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" - ".inst 0x4508a096 // sshllb z22.h, z4.b, #0x0\n" - ".inst 0x4508a494 // sshllt z20.h, z4.b, #0x0\n" - "ld1rw { z16.s }, p4/Z, [x20]\n" - ".inst 0x4510a261 // sshllb z1.s, z19.h, #0x0\n" - ".inst 0x4510a673 // sshllt z19.s, z19.h, #0x0\n" - ".inst 0x4510a220 // sshllb z0.s, z17.h, #0x0\n" - ".inst 0x4510a631 // sshllt z17.s, z17.h, #0x0\n" - ".inst 0x4510a25f // sshllb z31.s, z18.h, #0x0\n" - ".inst 0x4510a652 // sshllt z18.s, z18.h, #0x0\n" - ".inst 0x4510a31e // sshllb z30.s, z24.h, #0x0\n" - ".inst 0x4510a71d // sshllt z29.s, z24.h, #0x0\n" - ".inst 0x44829061 // srshl z1.s, p4/M, z1.s, z3.s\n" - ".inst 0x44829073 // srshl z19.s, p4/M, z19.s, z3.s\n" - ".inst 0x4510a2bc // sshllb z28.s, z21.h, #0x0\n" - ".inst 0x4510a6b5 // sshllt z21.s, z21.h, #0x0\n" - ".inst 0x44829060 // srshl z0.s, p4/M, z0.s, z3.s\n" - ".inst 0x44829071 // srshl z17.s, p4/M, z17.s, z3.s\n" - ".inst 0x4510a2fb // sshllb z27.s, z23.h, #0x0\n" - ".inst 0x4510a6fa // sshllt z26.s, z23.h, #0x0\n" + ".inst 0x4508a077 // sshllb z23.h, z3.b, #0\n" + ".inst 0x4508a476 // sshllt z22.h, z3.b, #0\n" + "ld1rw { z4.s }, p4/Z, [x21]\n" + "ld1rw { z3.s }, p4/Z, [x20]\n" + "neg z6.s, p4/M, z6.s\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" + "mov z2.s, #0x7f\n" + "ld1rw { z1.s }, p4/Z, [x20]\n" + "not z0.s, p4/M, z2.s\n" + ".inst 0x459340df // saddwb z31.s, z6.s, z19.h\n" + ".inst 0x459344d3 // saddwt z19.s, z6.s, z19.h\n" + ".inst 0x459140de // saddwb z30.s, z6.s, z17.h\n" + ".inst 0x459144d1 // saddwt z17.s, z6.s, z17.h\n" + ".inst 0x459240dd // saddwb z29.s, z6.s, z18.h\n" + ".inst 0x459244d2 // saddwt z18.s, z6.s, z18.h\n" + ".inst 0x459040dc // saddwb z28.s, z6.s, z16.h\n" + ".inst 0x459044d0 // saddwt z16.s, z6.s, z16.h\n" + ".inst 0x448290bf // srshl z31.s, p4/M, z31.s, z5.s\n" + ".inst 0x448290b3 // srshl z19.s, p4/M, z19.s, z5.s\n" + ".inst 0x459440d5 // saddwb z21.s, z6.s, z20.h\n" + ".inst 0x459444d4 // saddwt z20.s, z6.s, z20.h\n" + ".inst 0x448290be // srshl z30.s, p4/M, z30.s, z5.s\n" + ".inst 0x448290b1 // srshl z17.s, p4/M, z17.s, z5.s\n" + ".inst 0x459840db // saddwb z27.s, z6.s, z24.h\n" + ".inst 0x459844da // saddwt z26.s, z6.s, z24.h\n" + ".inst 0x448290bd // srshl z29.s, p4/M, z29.s, z5.s\n" + ".inst 0x448290b2 // srshl z18.s, p4/M, z18.s, z5.s\n" + ".inst 0x459740d9 // saddwb z25.s, z6.s, z23.h\n" + ".inst 0x459744d8 // saddwt z24.s, z6.s, z23.h\n" + ".inst 0x448290bc // srshl z28.s, p4/M, z28.s, z5.s\n" + ".inst 0x448290b0 // srshl z16.s, p4/M, z16.s, z5.s\n" + ".inst 0x459640d7 // saddwb z23.s, z6.s, z22.h\n" + ".inst 0x459644d6 // saddwt z22.s, z6.s, z22.h\n" + ".inst 0x448290b5 // srshl z21.s, p4/M, z21.s, z5.s\n" + ".inst 0x448290b4 // srshl z20.s, p4/M, z20.s, z5.s\n" + ".inst 0x448290bb // srshl z27.s, p4/M, z27.s, z5.s\n" + ".inst 0x448290ba // srshl z26.s, p4/M, z26.s, z5.s\n" + ".inst 0x04a477ff // sqrdmulh z31.s, z31.s, z4.s\n" + ".inst 0x04a47673 // sqrdmulh z19.s, z19.s, z4.s\n" + ".inst 0x448290b9 // srshl z25.s, p4/M, z25.s, z5.s\n" + ".inst 0x448290b8 // srshl z24.s, p4/M, z24.s, z5.s\n" + ".inst 0x04a477de // sqrdmulh z30.s, z30.s, z4.s\n" + ".inst 0x04a47631 // sqrdmulh z17.s, z17.s, z4.s\n" + ".inst 0x448290b7 // srshl z23.s, p4/M, z23.s, z5.s\n" + ".inst 0x448290b6 // srshl z22.s, p4/M, z22.s, z5.s\n" + ".inst 0x04a477bd // sqrdmulh z29.s, z29.s, z4.s\n" + ".inst 0x04a47652 // sqrdmulh z18.s, z18.s, z4.s\n" + ".inst 0x04a4779c // sqrdmulh z28.s, z28.s, z4.s\n" + ".inst 0x04a47610 // sqrdmulh z16.s, z16.s, z4.s\n" ".inst 0x4482907f // srshl z31.s, p4/M, z31.s, z3.s\n" - ".inst 0x44829072 // srshl z18.s, p4/M, z18.s, z3.s\n" - ".inst 0x4510a2d9 // sshllb z25.s, z22.h, #0x0\n" - ".inst 0x4510a6d8 // sshllt z24.s, z22.h, #0x0\n" + ".inst 0x44829073 // srshl z19.s, p4/M, z19.s, z3.s\n" + ".inst 0x04a476b5 // sqrdmulh z21.s, z21.s, z4.s\n" + ".inst 0x04a47694 // sqrdmulh z20.s, z20.s, z4.s\n" ".inst 0x4482907e // srshl z30.s, p4/M, z30.s, z3.s\n" + ".inst 0x44829071 // srshl z17.s, p4/M, z17.s, z3.s\n" + ".inst 0x04a4777b // sqrdmulh z27.s, z27.s, z4.s\n" + ".inst 0x04a4775a // sqrdmulh z26.s, z26.s, z4.s\n" ".inst 0x4482907d // srshl z29.s, p4/M, z29.s, z3.s\n" - ".inst 0x4510a297 // sshllb z23.s, z20.h, #0x0\n" - ".inst 0x4510a696 // sshllt z22.s, z20.h, #0x0\n" + ".inst 0x44829072 // srshl z18.s, p4/M, z18.s, z3.s\n" + ".inst 0x04a47739 // sqrdmulh z25.s, z25.s, z4.s\n" + ".inst 0x04a47718 // sqrdmulh z24.s, z24.s, z4.s\n" ".inst 0x4482907c // srshl z28.s, p4/M, z28.s, z3.s\n" + ".inst 0x44829070 // srshl z16.s, p4/M, z16.s, z3.s\n" + ".inst 0x04a476f7 // sqrdmulh z23.s, z23.s, z4.s\n" + ".inst 0x04a476d6 // sqrdmulh z22.s, z22.s, z4.s\n" ".inst 0x44829075 // srshl z21.s, p4/M, z21.s, z3.s\n" + ".inst 0x44829074 // srshl z20.s, p4/M, z20.s, z3.s\n" ".inst 0x4482907b // srshl z27.s, p4/M, z27.s, z3.s\n" ".inst 0x4482907a // srshl z26.s, p4/M, z26.s, z3.s\n" - ".inst 0x04a27421 // sqrdmulh z1.s, z1.s, z2.s\n" - ".inst 0x04a27673 // sqrdmulh z19.s, z19.s, z2.s\n" + "add z31.s, z31.s, z1.s\n" + "add z19.s, z19.s, z1.s\n" ".inst 0x44829079 // srshl z25.s, p4/M, z25.s, z3.s\n" ".inst 0x44829078 // srshl z24.s, p4/M, z24.s, z3.s\n" - ".inst 0x04a27400 // sqrdmulh z0.s, z0.s, z2.s\n" - ".inst 0x04a27631 // sqrdmulh z17.s, z17.s, z2.s\n" + "add z30.s, z30.s, z1.s\n" + "add z17.s, z17.s, z1.s\n" ".inst 0x44829077 // srshl z23.s, p4/M, z23.s, z3.s\n" ".inst 0x44829076 // srshl z22.s, p4/M, z22.s, z3.s\n" - ".inst 0x04a277ff // sqrdmulh z31.s, z31.s, z2.s\n" - ".inst 0x04a27652 // sqrdmulh z18.s, z18.s, z2.s\n" - ".inst 0x04a277de // sqrdmulh z30.s, z30.s, z2.s\n" - ".inst 0x04a277bd // sqrdmulh z29.s, z29.s, z2.s\n" - ".inst 0x44829201 // srshl z1.s, p4/M, z1.s, z16.s\n" - ".inst 0x44829213 // srshl z19.s, p4/M, z19.s, z16.s\n" - ".inst 0x04a2779c // sqrdmulh z28.s, z28.s, z2.s\n" - ".inst 0x04a276b5 // sqrdmulh z21.s, z21.s, z2.s\n" - ".inst 0x44829200 // srshl z0.s, p4/M, z0.s, z16.s\n" - ".inst 0x44829211 // srshl z17.s, p4/M, z17.s, z16.s\n" - ".inst 0x04a2777b // sqrdmulh z27.s, z27.s, z2.s\n" - ".inst 0x04a2775a // sqrdmulh z26.s, z26.s, z2.s\n" - ".inst 0x4482921f // srshl z31.s, p4/M, z31.s, z16.s\n" - ".inst 0x44829212 // srshl z18.s, p4/M, z18.s, z16.s\n" - ".inst 0x04a27739 // sqrdmulh z25.s, z25.s, z2.s\n" - ".inst 0x04a27718 // sqrdmulh z24.s, z24.s, z2.s\n" - ".inst 0x4482921e // srshl z30.s, p4/M, z30.s, z16.s\n" - ".inst 0x4482921d // srshl z29.s, p4/M, z29.s, z16.s\n" - ".inst 0x04a276f7 // sqrdmulh z23.s, z23.s, z2.s\n" - ".inst 0x04a276d6 // sqrdmulh z22.s, z22.s, z2.s\n" - ".inst 0x4482921c // srshl z28.s, p4/M, z28.s, z16.s\n" - ".inst 0x44829215 // srshl z21.s, p4/M, z21.s, z16.s\n" - "mov z20.s, #0x7f\n" - ".inst 0x4482921b // srshl z27.s, p4/M, z27.s, z16.s\n" - ".inst 0x4482921a // srshl z26.s, p4/M, z26.s, z16.s\n" - ".inst 0x44829219 // srshl z25.s, p4/M, z25.s, z16.s\n" - ".inst 0x44829218 // srshl z24.s, p4/M, z24.s, z16.s\n" - ".inst 0x44829217 // srshl z23.s, p4/M, z23.s, z16.s\n" - ".inst 0x44829216 // srshl z22.s, p4/M, z22.s, z16.s\n" - "not z16.s, p4/M, z20.s\n" - "smax z1.s, p4/M, z1.s, z16.s\n" - "smax z19.s, p4/M, z19.s, z16.s\n" - "smax z0.s, p4/M, z0.s, z16.s\n" - "smax z17.s, p4/M, z17.s, z16.s\n" - "smax z31.s, p4/M, z31.s, z16.s\n" - "smax z18.s, p4/M, z18.s, z16.s\n" - "smax z30.s, p4/M, z30.s, z16.s\n" - "smax z29.s, p4/M, z29.s, z16.s\n" - "smax z28.s, p4/M, z28.s, z16.s\n" - "smax z21.s, p4/M, z21.s, z16.s\n" - "smax z27.s, p4/M, z27.s, z16.s\n" - "smax z26.s, p4/M, z26.s, z16.s\n" - "smax z25.s, p4/M, z25.s, z16.s\n" - "smax z24.s, p4/M, z24.s, z16.s\n" - "smax z23.s, p4/M, z23.s, z16.s\n" - "smax z22.s, p4/M, z22.s, z16.s\n" - "smin z1.s, p4/M, z1.s, z20.s\n" - "smin z19.s, p4/M, z19.s, z20.s\n" - "smin z0.s, p4/M, z0.s, z20.s\n" - "smin z17.s, p4/M, z17.s, z20.s\n" - "smin z31.s, p4/M, z31.s, z20.s\n" - "smin z18.s, p4/M, z18.s, z20.s\n" - "smin z30.s, p4/M, z30.s, z20.s\n" - "smin z29.s, p4/M, z29.s, z20.s\n" - "smin z28.s, p4/M, z28.s, z20.s\n" - "trn1 z19.h, z1.h, z19.h\n" - "smin z21.s, p4/M, z21.s, z20.s\n" - "smin z27.s, p4/M, z27.s, z20.s\n" - "trn1 z17.h, z0.h, z17.h\n" - "smin z26.s, p4/M, z26.s, z20.s\n" - "smin z25.s, p4/M, z25.s, z20.s\n" - "trn1 z18.h, z31.h, z18.h\n" - "smin z24.s, p4/M, z24.s, z20.s\n" - "smin z23.s, p4/M, z23.s, z20.s\n" - "trn1 z16.h, z30.h, z29.h\n" - "smin z22.s, p4/M, z22.s, z20.s\n" - "trn1 z21.h, z28.h, z21.h\n" + "add z29.s, z29.s, z1.s\n" + "add z18.s, z18.s, z1.s\n" + "add z28.s, z28.s, z1.s\n" + "add z16.s, z16.s, z1.s\n" + "smax z31.s, p4/M, z31.s, z0.s\n" + "smax z19.s, p4/M, z19.s, z0.s\n" + "add z21.s, z21.s, z1.s\n" + "add z20.s, z20.s, z1.s\n" + "smax z30.s, p4/M, z30.s, z0.s\n" + "smax z17.s, p4/M, z17.s, z0.s\n" + "add z27.s, z27.s, z1.s\n" + "add z26.s, z26.s, z1.s\n" + "smax z29.s, p4/M, z29.s, z0.s\n" + "smax z18.s, p4/M, z18.s, z0.s\n" + "add z25.s, z25.s, z1.s\n" + "add z24.s, z24.s, z1.s\n" + "smax z28.s, p4/M, z28.s, z0.s\n" + "smax z16.s, p4/M, z16.s, z0.s\n" + "add z23.s, z23.s, z1.s\n" + "add z22.s, z22.s, z1.s\n" + "smax z21.s, p4/M, z21.s, z0.s\n" + "smax z20.s, p4/M, z20.s, z0.s\n" + "smax z27.s, p4/M, z27.s, z0.s\n" + "smax z26.s, p4/M, z26.s, z0.s\n" + "smax z25.s, p4/M, z25.s, z0.s\n" + "smax z24.s, p4/M, z24.s, z0.s\n" + "smax z23.s, p4/M, z23.s, z0.s\n" + "smax z22.s, p4/M, z22.s, z0.s\n" + "smin z31.s, p4/M, z31.s, z2.s\n" + "smin z19.s, p4/M, z19.s, z2.s\n" + "smin z30.s, p4/M, z30.s, z2.s\n" + "smin z17.s, p4/M, z17.s, z2.s\n" + "smin z29.s, p4/M, z29.s, z2.s\n" + "smin z18.s, p4/M, z18.s, z2.s\n" + "smin z28.s, p4/M, z28.s, z2.s\n" + "smin z16.s, p4/M, z16.s, z2.s\n" + "trn1 z19.h, z31.h, z19.h\n" + "smin z21.s, p4/M, z21.s, z2.s\n" + "smin z20.s, p4/M, z20.s, z2.s\n" + "trn1 z17.h, z30.h, z17.h\n" + "smin z27.s, p4/M, z27.s, z2.s\n" + "smin z26.s, p4/M, z26.s, z2.s\n" + "trn1 z18.h, z29.h, z18.h\n" + "smin z25.s, p4/M, z25.s, z2.s\n" + "smin z24.s, p4/M, z24.s, z2.s\n" + "trn1 z16.h, z28.h, z16.h\n" + "smin z23.s, p4/M, z23.s, z2.s\n" + "smin z22.s, p4/M, z22.s, z2.s\n" + "trn1 z21.h, z21.h, z20.h\n" "trn1 z20.b, z19.b, z17.b\n" "trn1 z17.h, z27.h, z26.h\n" "trn1 z19.h, z25.h, z24.h\n" @@ -285,16 +306,16 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "st1b { z16.b }, p0, [%x[outptr], x26]\n" "incb x26, ALL, MUL #4\n" "whilelt p0.b, x26, %x[n_channels]\n" - "b.any 1b\n" + "b.ne 1b\n" "7:" // Single vector of channels "whilelt p3.b, x9, %x[n_channels]\n" - "b.none 14f\n" + "b.eq 14f\n" "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" "mov z6.b, #0x80\n" "mov x24, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "subs x25, x25, #0x1\n" "add x24, x24, #0x20\n" @@ -306,7 +327,7 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "9:" // Single vector of channels: Loop: 4 inputs loop "movprfx z16, z2\n smax z16.b, p4/M, z16.b, z1.b\n" "movprfx z17, z23\n smax z17.b, p4/M, z17.b, z0.b\n" - "ldp x23, x22, [x24, #0x0]\n" + "ldp x23, x22, [x24, #0]\n" "ldp x21, x20, [x24, #0x10]\n" "subs x25, x25, #0x1\n" "add x24, x24, #0x20\n" @@ -332,50 +353,59 @@ void sve_s8q_nhwc_max_generic_depthfirst_impl( "smax z6.b, p4/M, z6.b, z16.b\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - ".inst 0x4508a0d1 // sshllb z17.h, z6.b, #0x0\n" - ".inst 0x4508a4d0 // sshllt z16.h, z6.b, #0x0\n" - "add x21, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x21, %x[quant_params], %[offsetof_qp_input_offset]\n" + ".inst 0x4508a0d1 // sshllb z17.h, z6.b, #0\n" + ".inst 0x4508a4da // sshllt z26.h, z6.b, #0\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z16.s }, p4/Z, [x21]\n" + "ld1rw { z25.s }, p4/Z, [x20]\n" + "add x21, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" "ld1rw { z24.s }, p4/Z, [x21]\n" "ld1rw { z23.s }, p4/Z, [x20]\n" - "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" "mov z22.s, #0x7f\n" "ld1rw { z21.s }, p4/Z, [x20]\n" - ".inst 0x4510a234 // sshllb z20.s, z17.h, #0x0\n" - ".inst 0x4510a631 // sshllt z17.s, z17.h, #0x0\n" - ".inst 0x4510a213 // sshllb z19.s, z16.h, #0x0\n" - ".inst 0x4510a612 // sshllt z18.s, z16.h, #0x0\n" - "not z16.s, p4/M, z22.s\n" - ".inst 0x44829314 // srshl z20.s, p4/M, z20.s, z24.s\n" - ".inst 0x44829311 // srshl z17.s, p4/M, z17.s, z24.s\n" - ".inst 0x44829313 // srshl z19.s, p4/M, z19.s, z24.s\n" - ".inst 0x44829312 // srshl z18.s, p4/M, z18.s, z24.s\n" - ".inst 0x04b77694 // sqrdmulh z20.s, z20.s, z23.s\n" - ".inst 0x04b77631 // sqrdmulh z17.s, z17.s, z23.s\n" - ".inst 0x04b77673 // sqrdmulh z19.s, z19.s, z23.s\n" - ".inst 0x04b77652 // sqrdmulh z18.s, z18.s, z23.s\n" - ".inst 0x448292b4 // srshl z20.s, p4/M, z20.s, z21.s\n" - ".inst 0x448292b1 // srshl z17.s, p4/M, z17.s, z21.s\n" - ".inst 0x448292b3 // srshl z19.s, p4/M, z19.s, z21.s\n" - ".inst 0x448292b2 // srshl z18.s, p4/M, z18.s, z21.s\n" - "smax z20.s, p4/M, z20.s, z16.s\n" - "smax z17.s, p4/M, z17.s, z16.s\n" - "smax z19.s, p4/M, z19.s, z16.s\n" - "smax z18.s, p4/M, z18.s, z16.s\n" - "smin z20.s, p4/M, z20.s, z22.s\n" - "smin z17.s, p4/M, z17.s, z22.s\n" + "neg z16.s, p4/M, z16.s\n" + "not z20.s, p4/M, z22.s\n" + ".inst 0x45914213 // saddwb z19.s, z16.s, z17.h\n" + ".inst 0x45914611 // saddwt z17.s, z16.s, z17.h\n" + ".inst 0x459a4212 // saddwb z18.s, z16.s, z26.h\n" + ".inst 0x459a4610 // saddwt z16.s, z16.s, z26.h\n" + ".inst 0x44829333 // srshl z19.s, p4/M, z19.s, z25.s\n" + ".inst 0x44829331 // srshl z17.s, p4/M, z17.s, z25.s\n" + ".inst 0x44829332 // srshl z18.s, p4/M, z18.s, z25.s\n" + ".inst 0x44829330 // srshl z16.s, p4/M, z16.s, z25.s\n" + ".inst 0x04b87673 // sqrdmulh z19.s, z19.s, z24.s\n" + ".inst 0x04b87631 // sqrdmulh z17.s, z17.s, z24.s\n" + ".inst 0x04b87652 // sqrdmulh z18.s, z18.s, z24.s\n" + ".inst 0x04b87610 // sqrdmulh z16.s, z16.s, z24.s\n" + ".inst 0x448292f3 // srshl z19.s, p4/M, z19.s, z23.s\n" + ".inst 0x448292f1 // srshl z17.s, p4/M, z17.s, z23.s\n" + ".inst 0x448292f2 // srshl z18.s, p4/M, z18.s, z23.s\n" + ".inst 0x448292f0 // srshl z16.s, p4/M, z16.s, z23.s\n" + "add z19.s, z19.s, z21.s\n" + "add z17.s, z17.s, z21.s\n" + "add z18.s, z18.s, z21.s\n" + "add z16.s, z16.s, z21.s\n" + "smax z19.s, p4/M, z19.s, z20.s\n" + "smax z17.s, p4/M, z17.s, z20.s\n" + "smax z18.s, p4/M, z18.s, z20.s\n" + "smax z16.s, p4/M, z16.s, z20.s\n" "smin z19.s, p4/M, z19.s, z22.s\n" + "smin z17.s, p4/M, z17.s, z22.s\n" "smin z18.s, p4/M, z18.s, z22.s\n" - "trn1 z17.h, z20.h, z17.h\n" - "trn1 z16.h, z19.h, z18.h\n" + "smin z16.s, p4/M, z16.s, z22.s\n" + "trn1 z17.h, z19.h, z17.h\n" + "trn1 z16.h, z18.h, z16.h\n" "trn1 z16.b, z17.b, z16.b\n" "st1b { z16.b }, p3, [%x[outptr], x9]\n" "incb x9\n" "whilelt p3.b, x9, %x[n_channels]\n" - "b.any 8b\n" + "b.ne 8b\n" "14:" // End : - : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [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)), [outptr] "r" (outptr), [quant_params] "r" (&qp) + : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [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)), [outptr] "r" (outptr), [quant_params] "r" (&qp) : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } diff --git a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp index fa1f5935bc..c3da624b6a 100644 --- a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp +++ b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp @@ -137,10 +137,6 @@ CpuPool2dAssemblyWrapperKernel::validate(const ITensorInfo *src, const ITensorIn ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.pool_type != PoolingType::MAX, "Assembly kernels only support differing src/dst quantization info for " "MAX pooling"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->data_type() == DataType::QASYMM8_SIGNED && - (src_qinfo.offset != 0 || dst_qinfo.offset != 0), - "Assembly kernels only support differing src/dst quantization info for " - "QASYMM8_SIGNED when both offsets are zero"); const float multiplier = src_qinfo.scale / dst_qinfo.scale; int32_t dst_multiplier{}; int32_t dst_shift{}; diff --git a/tests/validation/NEON/PoolingLayer.cpp b/tests/validation/NEON/PoolingLayer.cpp index 820a7e6264..6d2183b5ad 100644 --- a/tests/validation/NEON/PoolingLayer.cpp +++ b/tests/validation/NEON/PoolingLayer.cpp @@ -136,6 +136,17 @@ const auto PoolingDatasetQASYMM8SignedPaddedMaxDifferentQInfoZeroOffset = make("InputQuantInfo", {QuantizationInfo(0.25f, 0)}), make("OutputQuantInfo", {QuantizationInfo(0.5f, 0)})); +const auto PoolingDatasetQASYMM8SignedPaddedMaxDifferentQInfo = + combine(make("Shape", {TensorShape(7U, 5U, 3U)}), + make("PoolingType", {PoolingType::MAX}), + make("PoolingSize", {Size2D(3, 3)}), + make("PadStride", {PadStrideInfo(2, 2, 1, 1)}), + make("ExcludePadding", {false}), + make("DataType", DataType::QASYMM8_SIGNED), + make("DataLayout", {DataLayout::NHWC}), + make("InputQuantInfo", {QuantizationInfo(0.25f, -11)}), + make("OutputQuantInfo", {QuantizationInfo(0.5f, -7)})); + const auto PoolingDatasetQASYMM8SignedPaddedMaxSameQInfo = combine(make("Shape", {TensorShape(7U, 5U, 3U), TensorShape(8U, 7U, 5U)}), make("PoolingType", {PoolingType::MAX}), @@ -235,6 +246,38 @@ TEST_CASE(ValidatePaddedAvgDifferentQuantizationInfo, framework::DatasetMode::AL ARM_COMPUTE_EXPECT(!is_valid, framework::LogLevel::ERRORS); } +TEST_CASE(ValidatePaddedSignedMaxDifferentQuantizationInfo, framework::DatasetMode::ALL) +{ + TensorInfo input_info(TensorShape(3U, 15U, 11U, 1U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC); + TensorInfo output_info(TensorShape(3U, 8U, 6U, 1U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC); + const auto pool_info = PoolingLayerInfo(PoolingType::MAX, 3, DataLayout::NHWC, PadStrideInfo(2, 2, 1, 1), false); + + input_info.set_quantization_info(QuantizationInfo(0.25f, -11)); + output_info.set_quantization_info(QuantizationInfo(0.5f, -7)); + input_info.set_is_resizable(false); + output_info.set_is_resizable(false); + + const bool is_valid = bool(NEPoolingLayer::validate(&input_info, &output_info, pool_info)); + + ARM_COMPUTE_EXPECT(is_valid, framework::LogLevel::ERRORS); +} + +TEST_CASE(ValidatePaddedSignedAvgDifferentQuantizationInfo, framework::DatasetMode::ALL) +{ + TensorInfo input_info(TensorShape(3U, 15U, 11U, 1U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC); + TensorInfo output_info(TensorShape(3U, 8U, 6U, 1U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC); + const auto pool_info = PoolingLayerInfo(PoolingType::AVG, 3, DataLayout::NHWC, PadStrideInfo(2, 2, 1, 1), false); + + input_info.set_quantization_info(QuantizationInfo(0.25f, -11)); + output_info.set_quantization_info(QuantizationInfo(0.5f, -7)); + input_info.set_is_resizable(false); + output_info.set_is_resizable(false); + + const bool is_valid = bool(NEPoolingLayer::validate(&input_info, &output_info, pool_info)); + + ARM_COMPUTE_EXPECT(!is_valid, framework::LogLevel::ERRORS); +} + template using NEPoolingLayerIndicesFixture = PoolingLayerIndicesValidationFixture; @@ -538,6 +581,14 @@ FIXTURE_DATA_TEST_CASE(PaddedMaxDifferentQInfoZeroOffset, // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8_s); } +FIXTURE_DATA_TEST_CASE(PaddedMaxDifferentQInfo, + NEPoolingLayerQuantizedFixture, + framework::DatasetMode::PRECOMMIT, + PoolingDatasetQASYMM8SignedPaddedMaxDifferentQInfo) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_s); +} FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerQuantizedFixture, framework::DatasetMode::PRECOMMIT,