From d216f570750b8ccde3754c4aef53fc20a90cb32d Mon Sep 17 00:00:00 2001 From: Freddie Liardet Date: Tue, 3 Aug 2021 15:57:32 +0100 Subject: Update cpu depthwise kernels Resolves: COMPMID-4688 Signed-off-by: Freddie Liardet Change-Id: I9e22f967f5b7ccaebff2fc49f0253f621d62d820 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6030 Tested-by: Arm Jenkins Reviewed-by: Pablo Marquez Tello Reviewed-by: Georgios Pinitas Comments-Addressed: Arm Jenkins --- .../generic_direct.cpp | 200 ++++++++++----------- 1 file changed, 96 insertions(+), 104 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp index 2238bf08cd..d20e9913ae 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp @@ -25,7 +25,7 @@ #include #include -#if defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS) +#if __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS) namespace arm_conv { namespace depthwise { @@ -113,225 +113,217 @@ void sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_direct_impl( "add x9, x21, x22, LSL #1\n" "ld1rh { z18.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" "add x28, x9, x22, LSL #1\n" - "ld1h { z17.h }, p3/Z, [x16]\n" // Load from weights and bias - "mov z31.d, z17.d\n" - "ld1h { z0.h }, p3/Z, [x16, #1, MUL VL]\n" // Load from weights and bias + "ld1h { z17.h }, p3/Z, [x16]\n" "add x27, x28, x22, LSL #1\n" - "mov z30.d, z17.d\n" - "ld1h { z1.h }, p3/Z, [x16, #2, MUL VL]\n" // Load from weights and bias + "ld1h { z0.h }, p3/Z, [x16, #1, MUL VL]\n" "add x26, x27, x22, LSL #1\n" - "mov z29.d, z17.d\n" - "ld1h { z2.h }, p3/Z, [x16, #3, MUL VL]\n" // Load from weights and bias + "ld1h { z1.h }, p3/Z, [x16, #2, MUL VL]\n" "add x25, x13, x13\n" - "mov z28.d, z17.d\n" - "ld1h { z3.h }, p3/Z, [x16, #4, MUL VL]\n" // Load from weights and bias + "ld1h { z2.h }, p3/Z, [x16, #3, MUL VL]\n" "add x24, x25, x13\n" - "ld1h { z4.h }, p3/Z, [x16, #5, MUL VL]\n" // Load from weights and bias + "ld1h { z3.h }, p3/Z, [x16, #4, MUL VL]\n" "add x23, x24, x13\n" - "ld1h { z5.h }, p3/Z, [x16, #6, MUL VL]\n" // Load from weights and bias + "ld1h { z4.h }, p3/Z, [x16, #5, MUL VL]\n" "mul x19, x7, x20\n" // offset = tile_i * ld_output_row - "ld1h { z6.h }, p3/Z, [x16, #7, MUL VL]\n" // Load from weights and bias + "ld1h { z5.h }, p3/Z, [x16, #6, MUL VL]\n" "madd x19, x8, x11, x19\n" // offset += tile_j * ld_output_col - "whilelt p2.h, XZR, %x[n_channels]\n" - "ld1h { z9.h }, p2/Z, [x28, x25, LSL #1]\n" // Load input point (2, 2) - "ld1h { z10.h }, p2/Z, [x21]\n" // Load input point (0, 0) + "ld1h { z6.h }, p3/Z, [x16, #7, MUL VL]\n" "mul x19, x19, x17\n" // offset *= output_tile_size - "ld1h { z11.h }, p2/Z, [x21, x13, LSL #1]\n" // Load input point (0, 1) + "whilelt p2.h, XZR, %x[n_channels]\n" + "ld1h { z9.h }, p2/Z, [x28, x25, LSL #1]\n" + "ld1h { z10.h }, p2/Z, [x21]\n" "add x10, x10, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16) - "ld1h { z12.h }, p2/Z, [x21, x24, LSL #1]\n" // Load input point (0, 3) + "ld1h { z11.h }, p2/Z, [x21, x13, LSL #1]\n" "add x22, x10, x20, LSL #1\n" - "ld1h { z13.h }, p2/Z, [x21, x23, LSL #1]\n" // Load input point (0, 4) + "ld1h { z12.h }, p2/Z, [x21, x24, LSL #1]\n" "addvl x16, x16, #16\n" - "ld1h { z14.h }, p2/Z, [x9]\n" // Load input point (1, 0) + "ld1h { z13.h }, p2/Z, [x21, x23, LSL #1]\n" "cmp x14, %x[n_channels]\n" - "ld1h { z7.h }, p3/Z, [x16, #-8, MUL VL]\n" // Load from weights and bias - "ld1h { z8.h }, p3/Z, [x16, #-7, MUL VL]\n" // Load from weights and bias + "ld1h { z7.h }, p3/Z, [x16, #-8, MUL VL]\n" + "ld1h { z8.h }, p3/Z, [x16, #-7, MUL VL]\n" "addvl x16, x16, #-6\n" - "ld1h { z15.h }, p2/Z, [x9, x13, LSL #1]\n" // Load input point (1, 1) - "ld1h { z16.h }, p2/Z, [x21, x25, LSL #1]\n" // Load input point (0, 2) + "ld1h { z14.h }, p2/Z, [x9]\n" + "ld1h { z15.h }, p2/Z, [x9, x13, LSL #1]\n" + "ld1h { z16.h }, p2/Z, [x21, x25, LSL #1]\n" "bge 3f\n" "2:" // Tile loop: Channel loop - "fmla z31.h, p3/M, z8.h, z9.h\n" - "ld1h { z17.h }, p3/Z, [x16]\n" // Load from weights and bias + "movprfx z31, z17\n fmla z31.h, p3/M, z8.h, z9.h\n" "whilelt p1.h, x14, %x[n_channels]\n" - "fmla z30.h, p3/M, z6.h, z9.h\n" + "movprfx z30, z17\n fmla z30.h, p3/M, z6.h, z9.h\n" "inch x12\n" - "fmla z29.h, p3/M, z2.h, z9.h\n" + "movprfx z29, z17\n fmla z29.h, p3/M, z2.h, z9.h\n" "mov p0.b, p2.b\n" - "fmla z28.h, p3/M, z0.h, z9.h\n" + "movprfx z28, z17\n fmla z28.h, p3/M, z0.h, z9.h\n" + "ld1h { z17.h }, p3/Z, [x16]\n" "inch x15\n" - "addvl x21, x21, #1\n" "fmla z31.h, p3/M, z0.h, z10.h\n" - "ld1h { z10.h }, p1/Z, [x21]\n" // Load input point (0, 0) + "addvl x21, x21, #1\n" + "ld1h { z10.h }, p1/Z, [x21]\n" "fmla z30.h, p3/M, z1.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x9, x23, LSL #1]\n" // Load input point (1, 4) "inch x14\n" "fmla z31.h, p3/M, z1.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x9, x24, LSL #1]\n" // Load input point (1, 3) + "ld1h { z11.h }, p2/Z, [x9, x24, LSL #1]\n" + "ld1h { z12.h }, p2/Z, [x9, x23, LSL #1]\n" "fmla z30.h, p3/M, z2.h, z13.h\n" - "ld1h { z13.h }, p2/Z, [x9, x25, LSL #1]\n" // Load input point (1, 2) - "addvl x9, x9, #1\n" + "ld1h { z13.h }, p2/Z, [x9, x25, LSL #1]\n" "fmla z31.h, p3/M, z3.h, z14.h\n" - "ld1h { z14.h }, p2/Z, [x27]\n" // Load input point (3, 0) + "ld1h { z14.h }, p2/Z, [x27]\n" + "addvl x9, x9, #1\n" "fmla z30.h, p3/M, z0.h, z16.h\n" "fmla z29.h, p3/M, z3.h, z14.h\n" - "ld1h { z14.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (3, 4) + "ld1h { z14.h }, p2/Z, [x27, x23, LSL #1]\n" "fmla z31.h, p3/M, z4.h, z15.h\n" - "ld1h { z15.h }, p2/Z, [x28]\n" // Load input point (2, 0) + "ld1h { z15.h }, p2/Z, [x28]\n" "fmla z30.h, p3/M, z4.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x27, x13, LSL #1]\n" // Load input point (3, 1) + "ld1h { z11.h }, p2/Z, [x27, x13, LSL #1]\n" "fmla z29.h, p3/M, z0.h, z15.h\n" - "ld1h { z0.h }, p3/Z, [x16, #1, MUL VL]\n" // Load from weights and bias + "ld1h { z0.h }, p3/Z, [x16, #1, MUL VL]\n" "fmla z31.h, p3/M, z2.h, z16.h\n" - "ld1h { z16.h }, p2/Z, [x28, x13, LSL #1]\n" // Load input point (2, 1) + "ld1h { z16.h }, p2/Z, [x28, x13, LSL #1]\n" "fmla z30.h, p3/M, z5.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x28, x24, LSL #1]\n" // Load input point (2, 3) + "ld1h { z12.h }, p2/Z, [x28, x24, LSL #1]\n" "fmla z29.h, p3/M, z4.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x28, x23, LSL #1]\n" // Load input point (2, 4) + "ld1h { z11.h }, p2/Z, [x28, x23, LSL #1]\n" "addvl x28, x28, #1\n" "fmla z31.h, p3/M, z5.h, z13.h\n" - "ld1h { z9.h }, p1/Z, [x28, x25, LSL #1]\n" // Load input point (2, 2) + "ld1h { z9.h }, p1/Z, [x28, x25, LSL #1]\n" "fmla z30.h, p3/M, z3.h, z13.h\n" - "ld1h { z13.h }, p2/Z, [x27, x24, LSL #1]\n" // Load input point (3, 3) + "ld1h { z13.h }, p2/Z, [x27, x24, LSL #1]\n" "fmla z29.h, p3/M, z1.h, z16.h\n" "fmla z31.h, p3/M, z6.h, z15.h\n" - "ld1h { z15.h }, p2/Z, [x26]\n" // Load input point (4, 0) + "ld1h { z15.h }, p2/Z, [x26]\n" "fmla z28.h, p3/M, z4.h, z13.h\n" - "ld1h { z13.h }, p2/Z, [x26, x13, LSL #1]\n" // Load input point (4, 1) + "ld1h { z13.h }, p2/Z, [x26, x13, LSL #1]\n" "fmla z30.h, p3/M, z7.h, z12.h\n" - "ld1h { z4.h }, p3/Z, [x16, #5, MUL VL]\n" // Load from weights and bias + "ld1h { z4.h }, p3/Z, [x16, #5, MUL VL]\n" "fmla z29.h, p3/M, z6.h, z15.h\n" - "ld1h { z15.h }, p2/Z, [x26, x25, LSL #1]\n" // Load input point (4, 2) + "ld1h { z15.h }, p2/Z, [x26, x25, LSL #1]\n" "fmla z31.h, p3/M, z7.h, z16.h\n" - "ld1h { z16.h }, p2/Z, [x27, x25, LSL #1]\n" // Load input point (3, 2) + "ld1h { z16.h }, p2/Z, [x27, x25, LSL #1]\n" "addvl x27, x27, #1\n" "fmla z28.h, p3/M, z1.h, z12.h\n" - "ld1h { z12.h }, p1/Z, [x21, x24, LSL #1]\n" // Load input point (0, 3) + "ld1h { z12.h }, p1/Z, [x21, x24, LSL #1]\n" "fmla z30.h, p3/M, z8.h, z11.h\n" - "ld1h { z1.h }, p3/Z, [x16, #2, MUL VL]\n" // Load from weights and bias + "ld1h { z1.h }, p3/Z, [x16, #2, MUL VL]\n" "fmla z29.h, p3/M, z7.h, z13.h\n" - "ld1h { z13.h }, p1/Z, [x21, x23, LSL #1]\n" // Load input point (0, 4) + "ld1h { z13.h }, p1/Z, [x21, x23, LSL #1]\n" "fmax z31.h, p3/M, z31.h, z19.h\n" "fmla z28.h, p3/M, z5.h, z14.h\n" - "ld1h { z14.h }, p2/Z, [x26, x24, LSL #1]\n" // Load input point (4, 3) + "ld1h { z14.h }, p2/Z, [x26, x24, LSL #1]\n" "fmax z30.h, p3/M, z30.h, z19.h\n" "fmla z29.h, p3/M, z5.h, z16.h\n" - "ld1h { z5.h }, p3/Z, [x16, #6, MUL VL]\n" // Load from weights and bias + "ld1h { z5.h }, p3/Z, [x16, #6, MUL VL]\n" "fmin z31.h, p3/M, z31.h, z18.h\n" - "st1h { z31.h }, p0, [x10]\n" // Store output point (0, 0) - "mov z31.d, z17.d\n" + "st1h { z31.h }, p0, [x10]\n" "fmla z28.h, p3/M, z2.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x26, x23, LSL #1]\n" // Load input point (4, 4) - "whilelt p2.h, x15, %x[n_channels]\n" "fmla z29.h, p3/M, z8.h, z15.h\n" - "ld1h { z2.h }, p3/Z, [x16, #3, MUL VL]\n" // Load from weights and bias - "addvl x26, x26, #1\n" + "ld1h { z11.h }, p2/Z, [x26, x23, LSL #1]\n" + "whilelt p2.h, x15, %x[n_channels]\n" "fmin z30.h, p3/M, z30.h, z18.h\n" - "st1h { z30.h }, p0, [x10, x11, LSL #1]\n" // Store output point (0, 1) - "mov z30.d, z17.d\n" - "addvl x10, x10, #1\n" + "ld1h { z2.h }, p3/Z, [x16, #3, MUL VL]\n" + "addvl x26, x26, #1\n" "fmla z28.h, p3/M, z3.h, z16.h\n" - "ld1h { z16.h }, p1/Z, [x21, x25, LSL #1]\n" // Load input point (0, 2) + "ld1h { z16.h }, p1/Z, [x21, x25, LSL #1]\n" "cmp x14, %x[n_channels]\n" "fmax z29.h, p3/M, z29.h, z19.h\n" - "ld1h { z3.h }, p3/Z, [x16, #4, MUL VL]\n" // Load from weights and bias + "ld1h { z3.h }, p3/Z, [x16, #4, MUL VL]\n" + "st1h { z30.h }, p0, [x10, x11, LSL #1]\n" "fmla z28.h, p3/M, z7.h, z14.h\n" - "ld1h { z14.h }, p1/Z, [x9]\n" // Load input point (1, 0) + "ld1h { z14.h }, p1/Z, [x9]\n" + "addvl x10, x10, #1\n" "fmin z29.h, p3/M, z29.h, z18.h\n" - "st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0) - "mov z29.d, z17.d\n" + "st1h { z29.h }, p0, [x22]\n" "fmla z28.h, p3/M, z6.h, z15.h\n" - "ld1h { z15.h }, p1/Z, [x9, x13, LSL #1]\n" // Load input point (1, 1) - "ld1h { z6.h }, p3/Z, [x16, #7, MUL VL]\n" // Load from weights and bias + "ld1h { z15.h }, p1/Z, [x9, x13, LSL #1]\n" "fmla z28.h, p3/M, z8.h, z11.h\n" - "ld1h { z11.h }, p1/Z, [x21, x13, LSL #1]\n" // Load input point (0, 1) - "addvl x16, x16, #16\n" + "ld1h { z11.h }, p1/Z, [x21, x13, LSL #1]\n" + "ld1h { z6.h }, p3/Z, [x16, #7, MUL VL]\n" "fmax z28.h, p3/M, z28.h, z19.h\n" - "ld1h { z7.h }, p3/Z, [x16, #-8, MUL VL]\n" // Load from weights and bias - "ld1h { z8.h }, p3/Z, [x16, #-7, MUL VL]\n" // Load from weights and bias + "addvl x16, x16, #16\n" + "ld1h { z7.h }, p3/Z, [x16, #-8, MUL VL]\n" "fmin z28.h, p3/M, z28.h, z18.h\n" - "st1h { z28.h }, p0, [x22, x11, LSL #1]\n" // Store output point (1, 1) - "mov z28.d, z17.d\n" - "addvl x22, x22, #1\n" + "ld1h { z8.h }, p3/Z, [x16, #-7, MUL VL]\n" "addvl x16, x16, #-6\n" + "st1h { z28.h }, p0, [x22, x11, LSL #1]\n" + "addvl x22, x22, #1\n" "blt 2b\n" "3:" // Tile loop: Channel tail - "fmla z31.h, p3/M, z8.h, z9.h\n" + "movprfx z31, z17\n fmla z31.h, p3/M, z8.h, z9.h\n" "ldr x7, [%x[params_struct], %[offsetof_args_tile_i]]\n" "mov p0.b, p2.b\n" - "fmla z30.h, p3/M, z6.h, z9.h\n" + "movprfx z30, z17\n fmla z30.h, p3/M, z6.h, z9.h\n" "ldr x8, [%x[params_struct], %[offsetof_args_tile_j]]\n" "add x21, x7, #0x1\n" - "fmla z29.h, p3/M, z2.h, z9.h\n" + "movprfx z29, z17\n fmla z29.h, p3/M, z2.h, z9.h\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" - "fmla z28.h, p3/M, z0.h, z9.h\n" + "movprfx z28, z17\n fmla z28.h, p3/M, z0.h, z9.h\n" "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" "add x8, x8, #0x1\n" "fmla z31.h, p3/M, z0.h, z10.h\n" "cmp x8, x19\n" "fmla z30.h, p3/M, z1.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x9, x23, LSL #1]\n" // Load input point (1, 4) + "ld1h { z12.h }, p2/Z, [x9, x23, LSL #1]\n" "fmla z31.h, p3/M, z1.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x9, x24, LSL #1]\n" // Load input point (1, 3) + "ld1h { z11.h }, p2/Z, [x9, x24, LSL #1]\n" "csel x8, x8, XZR, LT\n" "fmla z30.h, p3/M, z2.h, z13.h\n" - "ld1h { z13.h }, p2/Z, [x9, x25, LSL #1]\n" // Load input point (1, 2) + "ld1h { z13.h }, p2/Z, [x9, x25, LSL #1]\n" "csel x7, x7, x21, LT\n" "fmla z31.h, p3/M, z3.h, z14.h\n" - "ld1h { z14.h }, p2/Z, [x27]\n" // Load input point (3, 0) + "ld1h { z14.h }, p2/Z, [x27]\n" "cmp x7, x20\n" "fmla z30.h, p3/M, z0.h, z16.h\n" "fmla z29.h, p3/M, z3.h, z14.h\n" - "ld1h { z14.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (3, 4) + "ld1h { z14.h }, p2/Z, [x27, x23, LSL #1]\n" "fmla z31.h, p3/M, z4.h, z15.h\n" - "ld1h { z15.h }, p2/Z, [x28]\n" // Load input point (2, 0) + "ld1h { z15.h }, p2/Z, [x28]\n" "fmla z30.h, p3/M, z4.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x27, x13, LSL #1]\n" // Load input point (3, 1) + "ld1h { z11.h }, p2/Z, [x27, x13, LSL #1]\n" "fmla z29.h, p3/M, z0.h, z15.h\n" "fmla z31.h, p3/M, z2.h, z16.h\n" - "ld1h { z16.h }, p2/Z, [x28, x13, LSL #1]\n" // Load input point (2, 1) + "ld1h { z16.h }, p2/Z, [x28, x13, LSL #1]\n" "fmla z30.h, p3/M, z5.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x28, x24, LSL #1]\n" // Load input point (2, 3) + "ld1h { z12.h }, p2/Z, [x28, x24, LSL #1]\n" "fmla z29.h, p3/M, z4.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x28, x23, LSL #1]\n" // Load input point (2, 4) + "ld1h { z11.h }, p2/Z, [x28, x23, LSL #1]\n" "fmla z31.h, p3/M, z5.h, z13.h\n" "fmla z30.h, p3/M, z3.h, z13.h\n" - "ld1h { z13.h }, p2/Z, [x27, x24, LSL #1]\n" // Load input point (3, 3) + "ld1h { z13.h }, p2/Z, [x27, x24, LSL #1]\n" "fmla z29.h, p3/M, z1.h, z16.h\n" "fmla z31.h, p3/M, z6.h, z15.h\n" - "ld1h { z15.h }, p2/Z, [x26]\n" // Load input point (4, 0) + "ld1h { z15.h }, p2/Z, [x26]\n" "fmla z28.h, p3/M, z4.h, z13.h\n" - "ld1h { z13.h }, p2/Z, [x26, x13, LSL #1]\n" // Load input point (4, 1) + "ld1h { z13.h }, p2/Z, [x26, x13, LSL #1]\n" "fmla z30.h, p3/M, z7.h, z12.h\n" "fmla z29.h, p3/M, z6.h, z15.h\n" - "ld1h { z15.h }, p2/Z, [x26, x25, LSL #1]\n" // Load input point (4, 2) + "ld1h { z15.h }, p2/Z, [x26, x25, LSL #1]\n" "fmla z31.h, p3/M, z7.h, z16.h\n" - "ld1h { z16.h }, p2/Z, [x27, x25, LSL #1]\n" // Load input point (3, 2) + "ld1h { z16.h }, p2/Z, [x27, x25, LSL #1]\n" "fmla z28.h, p3/M, z1.h, z12.h\n" "fmla z30.h, p3/M, z8.h, z11.h\n" "fmla z29.h, p3/M, z7.h, z13.h\n" "fmax z31.h, p3/M, z31.h, z19.h\n" "fmla z28.h, p3/M, z5.h, z14.h\n" - "ld1h { z14.h }, p2/Z, [x26, x24, LSL #1]\n" // Load input point (4, 3) + "ld1h { z14.h }, p2/Z, [x26, x24, LSL #1]\n" "fmax z30.h, p3/M, z30.h, z19.h\n" "fmla z29.h, p3/M, z5.h, z16.h\n" "fmin z31.h, p3/M, z31.h, z18.h\n" - "st1h { z31.h }, p0, [x10]\n" // Store output point (0, 0) + "st1h { z31.h }, p0, [x10]\n" "fmla z28.h, p3/M, z2.h, z11.h\n" "fmla z29.h, p3/M, z8.h, z15.h\n" - "ld1h { z11.h }, p2/Z, [x26, x23, LSL #1]\n" // Load input point (4, 4) + "ld1h { z11.h }, p2/Z, [x26, x23, LSL #1]\n" "fmin z30.h, p3/M, z30.h, z18.h\n" - "st1h { z30.h }, p0, [x10, x11, LSL #1]\n" // Store output point (0, 1) + "st1h { z30.h }, p0, [x10, x11, LSL #1]\n" "fmla z28.h, p3/M, z3.h, z16.h\n" "fmax z29.h, p3/M, z29.h, z19.h\n" "fmla z28.h, p3/M, z7.h, z14.h\n" "fmin z29.h, p3/M, z29.h, z18.h\n" - "st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0) + "st1h { z29.h }, p0, [x22]\n" "fmla z28.h, p3/M, z6.h, z15.h\n" "fmla z28.h, p3/M, z8.h, z11.h\n" "fmax z28.h, p3/M, z28.h, z19.h\n" "fmin z28.h, p3/M, z28.h, z18.h\n" - "st1h { z28.h }, p0, [x22, x11, LSL #1]\n" // Store output point (1, 1) + "st1h { z28.h }, p0, [x22, x11, LSL #1]\n" "blt 1b\n" : : [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (¶ms_struct) @@ -342,4 +334,4 @@ void sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_direct_impl( } // namespace depthwise } // namespace arm_conv -#endif // defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS) +#endif // __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) && defined(__ARM_FP16_ARGS) -- cgit v1.2.1