diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst')
2 files changed, 211 insertions, 207 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp index a4c1a40100..955a02de57 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp @@ -25,7 +25,7 @@ #include <cstddef> #include <cstdint> -#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 { @@ -113,149 +113,141 @@ void sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "add x27, x10, x22, LSL #1\n" "ld1rh { z17.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" "add x26, x27, x22, LSL #1\n" - "ld1h { z16.h }, p3/Z, [x14]\n" // Load from weights and bias - "mov z31.d, z16.d\n" - "ld1h { z0.h }, p3/Z, [x14, #1, MUL VL]\n" // Load from weights and bias + "ld1h { z16.h }, p3/Z, [x14]\n" "add x25, x26, x22, LSL #1\n" - "mov z30.d, z16.d\n" - "ld1h { z1.h }, p3/Z, [x14, #2, MUL VL]\n" // Load from weights and bias + "ld1h { z0.h }, p3/Z, [x14, #1, MUL VL]\n" "add x24, x11, x11\n" - "mov z29.d, z16.d\n" - "ld1h { z2.h }, p3/Z, [x14, #3, MUL VL]\n" // Load from weights and bias + "ld1h { z1.h }, p3/Z, [x14, #2, MUL VL]\n" "add x23, x24, x11\n" - "mov z28.d, z16.d\n" - "ld1h { z3.h }, p3/Z, [x14, #4, MUL VL]\n" // Load from weights and bias + "ld1h { z2.h }, p3/Z, [x14, #3, MUL VL]\n" "mul x19, x17, x20\n" // offset = tile_i * ld_output_row - "ld1h { z4.h }, p3/Z, [x14, #5, MUL VL]\n" // Load from weights and bias + "ld1h { z3.h }, p3/Z, [x14, #4, MUL VL]\n" "madd x19, x16, x9, x19\n" // offset += tile_j * ld_output_col - "ld1h { z5.h }, p3/Z, [x14, #6, MUL VL]\n" // Load from weights and bias + "ld1h { z4.h }, p3/Z, [x14, #5, MUL VL]\n" "mul x19, x19, x15\n" // offset *= output_tile_size - "ld1h { z6.h }, p3/Z, [x14, #7, MUL VL]\n" // Load from weights and bias + "ld1h { z5.h }, p3/Z, [x14, #6, MUL VL]\n" "add x28, x28, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16) - "whilelt p2.h, XZR, %x[n_channels]\n" - "ld1h { z9.h }, p2/Z, [x27, x11, LSL #1]\n" // Load input point (1, 1) - "ld1h { z10.h }, p2/Z, [x10]\n" // Load input point (0, 0) + "ld1h { z6.h }, p3/Z, [x14, #7, MUL VL]\n" "add x22, x28, x20, LSL #1\n" - "ld1h { z11.h }, p2/Z, [x10, x23, LSL #1]\n" // Load input point (0, 3) + "whilelt p2.h, XZR, %x[n_channels]\n" + "ld1h { z9.h }, p2/Z, [x27, x11, LSL #1]\n" + "ld1h { z10.h }, p2/Z, [x10]\n" "addvl x14, x14, #16\n" - "ld1h { z12.h }, p2/Z, [x27, x24, LSL #1]\n" // Load input point (1, 2) + "ld1h { z11.h }, p2/Z, [x10, x23, LSL #1]\n" "cmp x12, %x[n_channels]\n" - "ld1h { z7.h }, p3/Z, [x14, #-8, MUL VL]\n" // Load from weights and bias - "ld1h { z8.h }, p3/Z, [x14, #-7, MUL VL]\n" // Load from weights and bias + "ld1h { z7.h }, p3/Z, [x14, #-8, MUL VL]\n" + "ld1h { z8.h }, p3/Z, [x14, #-7, MUL VL]\n" "addvl x14, x14, #-6\n" - "ld1h { z13.h }, p2/Z, [x26, x11, LSL #1]\n" // Load input point (2, 1) + "ld1h { z12.h }, p2/Z, [x27, x24, LSL #1]\n" + "ld1h { z13.h }, p2/Z, [x26, x11, LSL #1]\n" "bge 3f\n" "2:" // Tile loop: Channel loop - "fmla z31.h, p3/M, z4.h, z9.h\n" - "ld1h { z16.h }, p3/Z, [x14]\n" // Load from weights and bias + "movprfx z31, z16\n fmla z31.h, p3/M, z4.h, z9.h\n" "whilelt p1.h, x12, %x[n_channels]\n" - "fmla z30.h, p3/M, z3.h, z9.h\n" + "movprfx z30, z16\n fmla z30.h, p3/M, z3.h, z9.h\n" "inch x21\n" - "fmla z29.h, p3/M, z1.h, z9.h\n" + "movprfx z29, z16\n fmla z29.h, p3/M, z1.h, z9.h\n" "mov p0.b, p2.b\n" - "fmla z28.h, p3/M, z0.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x25]\n" // Load input point (3, 0) + "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z9.h\n" + "ld1h { z9.h }, p2/Z, [x25]\n" "inch x13\n" "fmla z31.h, p3/M, z0.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x26, x24, LSL #1]\n" // Load input point (2, 2) + "ld1h { z10.h }, p2/Z, [x26, x24, LSL #1]\n" "inch x12\n" "fmla z30.h, p3/M, z2.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x25, x23, LSL #1]\n" // Load input point (3, 3) + "ld1h { z11.h }, p2/Z, [x25, x23, LSL #1]\n" "fmla z29.h, p3/M, z2.h, z12.h\n" + "ld1h { z16.h }, p3/Z, [x14]\n" "fmla z28.h, p3/M, z1.h, z12.h\n" "fmla z31.h, p3/M, z5.h, z12.h\n" "fmla z30.h, p3/M, z4.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x10, x11, LSL #1]\n" // Load input point (0, 1) + "ld1h { z12.h }, p2/Z, [x10, x11, LSL #1]\n" "fmla z29.h, p3/M, z6.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x10, x24, LSL #1]\n" // Load input point (0, 2) + "ld1h { z9.h }, p2/Z, [x10, x24, LSL #1]\n" "addvl x10, x10, #1\n" "fmla z28.h, p3/M, z3.h, z13.h\n" "fmla z31.h, p3/M, z7.h, z13.h\n" "fmla z30.h, p3/M, z6.h, z13.h\n" "fmla z29.h, p3/M, z4.h, z13.h\n" "fmla z28.h, p3/M, z8.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x27]\n" // Load input point (1, 0) + "ld1h { z11.h }, p2/Z, [x27]\n" "fmla z31.h, p3/M, z1.h, z12.h\n" "fmla z30.h, p3/M, z0.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (1, 3) + "ld1h { z12.h }, p2/Z, [x27, x23, LSL #1]\n" "addvl x27, x27, #1\n" "fmla z29.h, p3/M, z5.h, z10.h\n" "fmla z28.h, p3/M, z4.h, z10.h\n" - "ld1h { z4.h }, p3/Z, [x14, #5, MUL VL]\n" // Load from weights and bias + "ld1h { z4.h }, p3/Z, [x14, #5, MUL VL]\n" "fmla z31.h, p3/M, z2.h, z9.h\n" "fmla z30.h, p3/M, z1.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x26]\n" // Load input point (2, 0) - "ld1h { z1.h }, p3/Z, [x14, #2, MUL VL]\n" // Load from weights and bias + "ld1h { z9.h }, p2/Z, [x26]\n" + "ld1h { z1.h }, p3/Z, [x14, #2, MUL VL]\n" "fmla z29.h, p3/M, z0.h, z11.h\n" - "ld1h { z0.h }, p3/Z, [x14, #1, MUL VL]\n" // Load from weights and bias + "ld1h { z0.h }, p3/Z, [x14, #1, MUL VL]\n" "fmla z28.h, p3/M, z2.h, z12.h\n" - "ld1h { z2.h }, p3/Z, [x14, #3, MUL VL]\n" // Load from weights and bias + "ld1h { z2.h }, p3/Z, [x14, #3, MUL VL]\n" "fmla z31.h, p3/M, z8.h, z10.h\n" "fmla z30.h, p3/M, z7.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x26, x23, LSL #1]\n" // Load input point (2, 3) + "ld1h { z10.h }, p2/Z, [x26, x23, LSL #1]\n" "addvl x26, x26, #1\n" "fmla z29.h, p3/M, z3.h, z9.h\n" - "ld1h { z13.h }, p1/Z, [x26, x11, LSL #1]\n" // Load input point (2, 1) + "ld1h { z13.h }, p1/Z, [x26, x11, LSL #1]\n" "fmla z31.h, p3/M, z3.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x25, x11, LSL #1]\n" // Load input point (3, 1) + "ld1h { z11.h }, p2/Z, [x25, x11, LSL #1]\n" "fmla z28.h, p3/M, z5.h, z10.h\n" - "ld1h { z3.h }, p3/Z, [x14, #4, MUL VL]\n" // Load from weights and bias + "ld1h { z3.h }, p3/Z, [x14, #4, MUL VL]\n" "fmla z30.h, p3/M, z5.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x25, x24, LSL #1]\n" // Load input point (3, 2) + "ld1h { z12.h }, p2/Z, [x25, x24, LSL #1]\n" "whilelt p2.h, x13, %x[n_channels]\n" "fmla z29.h, p3/M, z7.h, z11.h\n" - "ld1h { z5.h }, p3/Z, [x14, #6, MUL VL]\n" // Load from weights and bias + "ld1h { z5.h }, p3/Z, [x14, #6, MUL VL]\n" "addvl x25, x25, #1\n" "fmla z31.h, p3/M, z6.h, z9.h\n" - "ld1h { z9.h }, p1/Z, [x27, x11, LSL #1]\n" // Load input point (1, 1) + "ld1h { z9.h }, p1/Z, [x27, x11, LSL #1]\n" "cmp x12, %x[n_channels]\n" "fmla z30.h, p3/M, z8.h, z10.h\n" - "ld1h { z10.h }, p1/Z, [x10]\n" // Load input point (0, 0) + "ld1h { z10.h }, p1/Z, [x10]\n" "fmla z28.h, p3/M, z6.h, z11.h\n" - "ld1h { z11.h }, p1/Z, [x10, x23, LSL #1]\n" // Load input point (0, 3) - "ld1h { z6.h }, p3/Z, [x14, #7, MUL VL]\n" // Load from weights and bias + "ld1h { z11.h }, p1/Z, [x10, x23, LSL #1]\n" + "ld1h { z6.h }, p3/Z, [x14, #7, MUL VL]\n" "fmla z29.h, p3/M, z8.h, z12.h\n" "addvl x14, x14, #16\n" "fmax z31.h, p3/M, z31.h, z18.h\n" - "ld1h { z8.h }, p3/Z, [x14, #-7, MUL VL]\n" // Load from weights and bias + "ld1h { z8.h }, p3/Z, [x14, #-7, MUL VL]\n" "fmla z28.h, p3/M, z7.h, z12.h\n" - "ld1h { z12.h }, p1/Z, [x27, x24, LSL #1]\n" // Load input point (1, 2) + "ld1h { z12.h }, p1/Z, [x27, x24, LSL #1]\n" "fmax z30.h, p3/M, z30.h, z18.h\n" - "ld1h { z7.h }, p3/Z, [x14, #-8, MUL VL]\n" // Load from weights and bias + "ld1h { z7.h }, p3/Z, [x14, #-8, MUL VL]\n" "addvl x14, x14, #-6\n" "fmax z29.h, p3/M, z29.h, z18.h\n" "fmin z31.h, p3/M, z31.h, z17.h\n" - "st1h { z31.h }, p0, [x28]\n" // Store output point (0, 0) - "mov z31.d, z16.d\n" + "st1h { z31.h }, p0, [x28]\n" "fmin z30.h, p3/M, z30.h, z17.h\n" - "st1h { z30.h }, p0, [x28, x9, LSL #1]\n" // Store output point (0, 1) - "mov z30.d, z16.d\n" - "addvl x28, x28, #1\n" "fmin z29.h, p3/M, z29.h, z17.h\n" - "st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0) - "mov z29.d, z16.d\n" + "st1h { z30.h }, p0, [x28, x9, LSL #1]\n" "fmax z28.h, p3/M, z28.h, z18.h\n" + "addvl x28, x28, #1\n" "fmin z28.h, p3/M, z28.h, z17.h\n" - "st1h { z28.h }, p0, [x22, x9, LSL #1]\n" // Store output point (1, 1) - "mov z28.d, z16.d\n" + "st1h { z29.h }, p0, [x22]\n" + "st1h { z28.h }, p0, [x22, x9, LSL #1]\n" "addvl x22, x22, #1\n" "blt 2b\n" "3:" // Tile loop: Channel tail - "fmla z31.h, p3/M, z4.h, z9.h\n" + "movprfx z31, z16\n fmla z31.h, p3/M, z4.h, z9.h\n" "ldr x17, [%x[params_struct], %[offsetof_args_tile_i]]\n" "mov p0.b, p2.b\n" - "fmla z30.h, p3/M, z3.h, z9.h\n" + "movprfx z30, z16\n fmla z30.h, p3/M, z3.h, z9.h\n" "ldr x16, [%x[params_struct], %[offsetof_args_tile_j]]\n" "add x21, x17, #0x1\n" - "fmla z29.h, p3/M, z1.h, z9.h\n" + "movprfx z29, z16\n fmla z29.h, p3/M, z1.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" - "ld1h { z9.h }, p2/Z, [x25]\n" // Load input point (3, 0) + "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z9.h\n" + "ld1h { z9.h }, p2/Z, [x25]\n" "add x16, x16, #0x1\n" "fmla z31.h, p3/M, z0.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x26, x24, LSL #1]\n" // Load input point (2, 2) + "ld1h { z10.h }, p2/Z, [x26, x24, LSL #1]\n" "fmla z30.h, p3/M, z2.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x25, x23, LSL #1]\n" // Load input point (3, 3) + "ld1h { z11.h }, p2/Z, [x25, x23, LSL #1]\n" "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" "fmla z29.h, p3/M, z2.h, z12.h\n" "cmp x16, x19\n" @@ -263,10 +255,10 @@ void sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "fmla z30.h, p3/M, z4.h, z12.h\n" "csel x16, x16, XZR, LT\n" "fmla z28.h, p3/M, z1.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x10, x11, LSL #1]\n" // Load input point (0, 1) + "ld1h { z12.h }, p2/Z, [x10, x11, LSL #1]\n" "csel x17, x17, x21, LT\n" "fmla z29.h, p3/M, z6.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x10, x24, LSL #1]\n" // Load input point (0, 2) + "ld1h { z9.h }, p2/Z, [x10, x24, LSL #1]\n" "cmp x17, x20\n" "fmla z31.h, p3/M, z7.h, z13.h\n" "fmla z30.h, p3/M, z6.h, z13.h\n" @@ -274,24 +266,24 @@ void sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "fmla z29.h, p3/M, z4.h, z13.h\n" "fmla z31.h, p3/M, z1.h, z12.h\n" "fmla z30.h, p3/M, z0.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (1, 3) + "ld1h { z12.h }, p2/Z, [x27, x23, LSL #1]\n" "fmla z28.h, p3/M, z8.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x27]\n" // Load input point (1, 0) + "ld1h { z11.h }, p2/Z, [x27]\n" "fmla z29.h, p3/M, z5.h, z10.h\n" "fmla z31.h, p3/M, z2.h, z9.h\n" "fmla z30.h, p3/M, z1.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x26]\n" // Load input point (2, 0) + "ld1h { z9.h }, p2/Z, [x26]\n" "fmla z28.h, p3/M, z4.h, z10.h\n" "fmla z29.h, p3/M, z0.h, z11.h\n" "fmla z31.h, p3/M, z8.h, z10.h\n" "fmla z30.h, p3/M, z7.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x26, x23, LSL #1]\n" // Load input point (2, 3) + "ld1h { z10.h }, p2/Z, [x26, x23, LSL #1]\n" "fmla z28.h, p3/M, z2.h, z12.h\n" "fmla z29.h, p3/M, z3.h, z9.h\n" "fmla z31.h, p3/M, z3.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x25, x11, LSL #1]\n" // Load input point (3, 1) + "ld1h { z11.h }, p2/Z, [x25, x11, LSL #1]\n" "fmla z30.h, p3/M, z5.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x25, x24, LSL #1]\n" // Load input point (3, 2) + "ld1h { z12.h }, p2/Z, [x25, x24, LSL #1]\n" "fmla z28.h, p3/M, z5.h, z10.h\n" "fmla z29.h, p3/M, z7.h, z11.h\n" "fmla z31.h, p3/M, z6.h, z9.h\n" @@ -303,14 +295,14 @@ void sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( "fmla z28.h, p3/M, z7.h, z12.h\n" "fmax z29.h, p3/M, z29.h, z18.h\n" "fmin z31.h, p3/M, z31.h, z17.h\n" - "st1h { z31.h }, p0, [x28]\n" // Store output point (0, 0) + "st1h { z31.h }, p0, [x28]\n" "fmin z30.h, p3/M, z30.h, z17.h\n" "fmin z29.h, p3/M, z29.h, z17.h\n" - "st1h { z30.h }, p0, [x28, x9, LSL #1]\n" // Store output point (0, 1) + "st1h { z30.h }, p0, [x28, x9, LSL #1]\n" "fmax z28.h, p3/M, z28.h, z18.h\n" - "st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0) + "st1h { z29.h }, p0, [x22]\n" "fmin z28.h, p3/M, z28.h, z17.h\n" - "st1h { z28.h }, p0, [x22, x9, LSL #1]\n" // Store output point (1, 1) + "st1h { z28.h }, p0, [x22, x9, 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) @@ -321,4 +313,4 @@ void sve_fp16_nhwc_3x3_s1_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) diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp index a845e7c0c6..7cca6fbcbf 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp @@ -25,7 +25,7 @@ #include <cstddef> #include <cstdint> -#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 { @@ -54,22 +54,22 @@ void sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_indirect_impl( const __fp16 max ) : outptrs(outptrs), params(params), min(min), max(max) { - inptrs[0] = input_ptrs[0]; - inptrs[1] = input_ptrs[1]; - inptrs[2] = input_ptrs[2]; - inptrs[3] = input_ptrs[3]; - inptrs[4] = input_ptrs[4]; - inptrs[5] = input_ptrs[5]; - inptrs[6] = input_ptrs[6]; - inptrs[7] = input_ptrs[7]; - inptrs[8] = input_ptrs[8]; - inptrs[9] = input_ptrs[9]; - inptrs[10] = input_ptrs[10]; - inptrs[11] = input_ptrs[11]; - inptrs[12] = input_ptrs[12]; - inptrs[13] = input_ptrs[13]; - inptrs[14] = input_ptrs[14]; - inptrs[15] = input_ptrs[15]; + inptrs[0] = input_ptrs[5]; + inptrs[1] = input_ptrs[0]; + inptrs[2] = input_ptrs[3]; + inptrs[3] = input_ptrs[6]; + inptrs[4] = input_ptrs[9]; + inptrs[5] = input_ptrs[12]; + inptrs[6] = input_ptrs[15]; + inptrs[7] = input_ptrs[1]; + inptrs[8] = input_ptrs[2]; + inptrs[9] = input_ptrs[10]; + inptrs[10] = input_ptrs[4]; + inptrs[11] = input_ptrs[7]; + inptrs[12] = input_ptrs[8]; + inptrs[13] = input_ptrs[11]; + inptrs[14] = input_ptrs[13]; + inptrs[15] = input_ptrs[14]; } }; @@ -78,207 +78,219 @@ void sve_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_indirect_impl( activation_min, activation_max); __asm__ __volatile__( - "ldr x3, [%x[params_struct], %[offsetof_args_outptrs]]\n" + "ldr x19, [%x[params_struct], %[offsetof_args_outptrs]]\n" "ptrue p3.b\n" - "ldr x4, [%x[params_struct], %[offsetof_args_params]]\n" - "add x19, %x[params_struct], %[offsetof_Args_inptrs]\n" + "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n" + "add x14, %x[params_struct], %[offsetof_Args_inptrs]\n" "ld1rh { z18.h }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" - "mov x5, #0x0\n" - "ldp x6, x7, [x19, #0x0]\n" - "cnth x8\n" - "ldp x17, x16, [x19, #0x10]\n" - "sub x15, XZR, x8\n" - "ldp x14, x13, [x19, #0x20]\n" - "whilelt p2.h, XZR, %x[n_channels]\n" - "ldp x12, x11, [x19, #0x30]\n" - "cmp x8, %x[n_channels]\n" - "ldp x10, x9, [x19, #0x40]\n" - "ldp x28, x27, [x19, #0x50]\n" - "ldp x26, x25, [x19, #0x60]\n" - "ldp x24, x23, [x19, #0x70]\n" - "ldp x22, x21, [x3, #0x0]\n" - "ldp x20, x19, [x3, #0x10]\n" + "mov x13, #0x0\n" "ld1rh { z17.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" - "ld1h { z16.h }, p3/Z, [x4]\n" // Load from weights and bias - "mov z31.d, z16.d\n" - "ld1h { z0.h }, p3/Z, [x4, #1, MUL VL]\n" // Load from weights and bias - "mov z30.d, z16.d\n" - "ld1h { z1.h }, p3/Z, [x4, #2, MUL VL]\n" // Load from weights and bias - "mov z29.d, z16.d\n" - "ld1h { z2.h }, p3/Z, [x4, #3, MUL VL]\n" // Load from weights and bias - "mov z28.d, z16.d\n" - "ld1h { z3.h }, p3/Z, [x4, #4, MUL VL]\n" // Load from weights and bias - "ld1h { z4.h }, p3/Z, [x4, #5, MUL VL]\n" // Load from weights and bias - "ld1h { z5.h }, p3/Z, [x4, #6, MUL VL]\n" // Load from weights and bias - "ld1h { z6.h }, p3/Z, [x4, #7, MUL VL]\n" // Load from weights and bias - "addvl x4, x4, #16\n" - "ld1h { z9.h }, p2/Z, [x13, x5, LSL #1]\n" - "ld1h { z7.h }, p3/Z, [x4, #-8, MUL VL]\n" // Load from weights and bias - "ld1h { z8.h }, p3/Z, [x4, #-7, MUL VL]\n" // Load from weights and bias - "addvl x4, x4, #-6\n" - "ld1h { z10.h }, p2/Z, [x6, x5, LSL #1]\n" - "ld1h { z11.h }, p2/Z, [x16, x5, LSL #1]\n" - "ld1h { z12.h }, p2/Z, [x12, x5, LSL #1]\n" - "ld1h { z13.h }, p2/Z, [x9, x5, LSL #1]\n" + "cnth x12\n" + "ldp x11, x10, [x19, #0x0]\n" + "sub x9, XZR, x12\n" + "ldp x28, x27, [x19, #0x10]\n" + "whilelt p2.h, XZR, %x[n_channels]\n" + "ld1h { z16.h }, p3/Z, [x15]\n" + "cmp x12, %x[n_channels]\n" + "ld1h { z0.h }, p3/Z, [x15, #1, MUL VL]\n" + "ld1h { z1.h }, p3/Z, [x15, #2, MUL VL]\n" + "ld1h { z2.h }, p3/Z, [x15, #3, MUL VL]\n" + "ld1h { z3.h }, p3/Z, [x15, #4, MUL VL]\n" + "ld1h { z4.h }, p3/Z, [x15, #5, MUL VL]\n" + "ld1h { z5.h }, p3/Z, [x15, #6, MUL VL]\n" + "ld1h { z6.h }, p3/Z, [x15, #7, MUL VL]\n" + "addvl x15, x15, #16\n" + "ldp x26, x25, [x14, #0x0]\n" + "ld1h { z7.h }, p3/Z, [x15, #-8, MUL VL]\n" + "ld1h { z8.h }, p3/Z, [x15, #-7, MUL VL]\n" + "addvl x15, x15, #-6\n" + "ld1h { z9.h }, p2/Z, [x26, x13, LSL #1]\n" + "ld1h { z10.h }, p2/Z, [x25, x13, LSL #1]\n" + "ldp x24, x23, [x14, #0x10]\n" + "ldr x22, [x14, #0x20]\n" + "ld1h { z11.h }, p2/Z, [x24, x13, LSL #1]\n" + "ld1h { z12.h }, p2/Z, [x23, x13, LSL #1]\n" + "ld1h { z13.h }, p2/Z, [x22, x13, LSL #1]\n" "bge 2f\n" "1:" // Channel loop - "fmla z31.h, p3/M, z4.h, z9.h\n" - "ld1h { z16.h }, p3/Z, [x4]\n" // Load from weights and bias - "whilelt p1.h, x8, %x[n_channels]\n" - "fmla z30.h, p3/M, z3.h, z9.h\n" - "inch x15\n" - "fmla z29.h, p3/M, z1.h, z9.h\n" + "movprfx z31, z16\n fmla z31.h, p3/M, z4.h, z9.h\n" + "ldr x21, [x14, #0x28]\n" + "whilelt p1.h, x12, %x[n_channels]\n" + "movprfx z30, z16\n fmla z30.h, p3/M, z3.h, z9.h\n" + "ldr x20, [x14, #0x30]\n" + "inch x9\n" + "movprfx z29, z16\n fmla z29.h, p3/M, z1.h, z9.h\n" + "ldr x19, [x14, #0x38]\n" "mov p0.b, p2.b\n" - "fmla z28.h, p3/M, z0.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x26, x5, LSL #1]\n" + "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z9.h\n" + "ld1h { z9.h }, p2/Z, [x21, x13, LSL #1]\n" + "ldr x26, [x14, #0x40]\n" "fmla z31.h, p3/M, z0.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x28, x5, LSL #1]\n" + "ldr x25, [x14, #0x48]\n" "fmla z30.h, p3/M, z2.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x23, x5, LSL #1]\n" + "ld1h { z11.h }, p2/Z, [x20, x13, LSL #1]\n" "fmla z29.h, p3/M, z2.h, z12.h\n" + "ldr x24, [x14, #0x50]\n" "fmla z28.h, p3/M, z1.h, z12.h\n" + "ld1h { z10.h }, p2/Z, [x25, x13, LSL #1]\n" "fmla z31.h, p3/M, z5.h, z12.h\n" + "ldr x23, [x14, #0x58]\n" "fmla z30.h, p3/M, z4.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x7, x5, LSL #1]\n" + "ld1h { z12.h }, p2/Z, [x19, x13, LSL #1]\n" "fmla z29.h, p3/M, z6.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x17, x5, LSL #1]\n" + "ld1h { z9.h }, p2/Z, [x26, x13, LSL #1]\n" "fmla z28.h, p3/M, z3.h, z13.h\n" + "ldr x22, [x14, #0x60]\n" "fmla z31.h, p3/M, z7.h, z13.h\n" + "ldr x21, [x14, #0x68]\n" "fmla z30.h, p3/M, z6.h, z13.h\n" + "ldr x20, [x14, #0x70]\n" "fmla z29.h, p3/M, z4.h, z13.h\n" - "ld1h { z13.h }, p1/Z, [x9, x8, LSL #1]\n" + "ldr x19, [x14, #0x78]\n" "fmla z28.h, p3/M, z8.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x14, x5, LSL #1]\n" + "ld1h { z11.h }, p2/Z, [x24, x13, LSL #1]\n" "fmla z31.h, p3/M, z1.h, z12.h\n" + "ldp x26, x25, [x14, #0x0]\n" "fmla z30.h, p3/M, z0.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x11, x5, LSL #1]\n" + "ld1h { z12.h }, p2/Z, [x23, x13, LSL #1]\n" "fmla z29.h, p3/M, z5.h, z10.h\n" + "ldp x24, x23, [x14, #0x10]\n" "fmla z28.h, p3/M, z4.h, z10.h\n" - "ld1h { z4.h }, p3/Z, [x4, #5, MUL VL]\n" // Load from weights and bias + "ld1h { z16.h }, p3/Z, [x15]\n" "fmla z31.h, p3/M, z2.h, z9.h\n" + "ld1h { z4.h }, p3/Z, [x15, #5, MUL VL]\n" "fmla z30.h, p3/M, z1.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x10, x5, LSL #1]\n" + "ld1h { z9.h }, p2/Z, [x22, x13, LSL #1]\n" "fmla z29.h, p3/M, z0.h, z11.h\n" - "ld1h { z0.h }, p3/Z, [x4, #1, MUL VL]\n" // Load from weights and bias + "ldr x22, [x14, #0x20]\n" "fmla z28.h, p3/M, z2.h, z12.h\n" - "ld1h { z1.h }, p3/Z, [x4, #2, MUL VL]\n" // Load from weights and bias - "ld1h { z2.h }, p3/Z, [x4, #3, MUL VL]\n" // Load from weights and bias + "ld1h { z0.h }, p3/Z, [x15, #1, MUL VL]\n" "fmla z31.h, p3/M, z8.h, z10.h\n" + "ld1h { z1.h }, p3/Z, [x15, #2, MUL VL]\n" "fmla z30.h, p3/M, z7.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x27, x5, LSL #1]\n" + "ld1h { z10.h }, p2/Z, [x21, x13, LSL #1]\n" "fmla z29.h, p3/M, z3.h, z9.h\n" + "ld1h { z13.h }, p1/Z, [x22, x12, LSL #1]\n" "fmla z31.h, p3/M, z3.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x25, x5, LSL #1]\n" + "ld1h { z11.h }, p2/Z, [x20, x13, LSL #1]\n" "fmla z28.h, p3/M, z5.h, z10.h\n" - "ld1h { z3.h }, p3/Z, [x4, #4, MUL VL]\n" // Load from weights and bias + "ld1h { z2.h }, p3/Z, [x15, #3, MUL VL]\n" "fmla z30.h, p3/M, z5.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x24, x5, LSL #1]\n" - "inch x5\n" + "ld1h { z12.h }, p2/Z, [x19, x13, LSL #1]\n" + "inch x13\n" "fmla z29.h, p3/M, z7.h, z11.h\n" - "ld1h { z5.h }, p3/Z, [x4, #6, MUL VL]\n" // Load from weights and bias - "whilelt p2.h, x5, %x[n_channels]\n" + "ld1h { z3.h }, p3/Z, [x15, #4, MUL VL]\n" + "whilelt p2.h, x13, %x[n_channels]\n" "fmla z31.h, p3/M, z6.h, z9.h\n" - "ld1h { z9.h }, p1/Z, [x13, x8, LSL #1]\n" + "ld1h { z9.h }, p1/Z, [x26, x12, LSL #1]\n" "fmla z28.h, p3/M, z6.h, z11.h\n" - "ld1h { z11.h }, p1/Z, [x16, x8, LSL #1]\n" + "ld1h { z11.h }, p1/Z, [x24, x12, LSL #1]\n" "fmla z30.h, p3/M, z8.h, z10.h\n" - "ld1h { z10.h }, p1/Z, [x6, x8, LSL #1]\n" - "ld1h { z6.h }, p3/Z, [x4, #7, MUL VL]\n" // Load from weights and bias + "ld1h { z10.h }, p1/Z, [x25, x12, LSL #1]\n" + "ld1h { z5.h }, p3/Z, [x15, #6, MUL VL]\n" "fmla z29.h, p3/M, z8.h, z12.h\n" - "addvl x4, x4, #16\n" + "ld1h { z6.h }, p3/Z, [x15, #7, MUL VL]\n" "fmla z28.h, p3/M, z7.h, z12.h\n" - "ld1h { z12.h }, p1/Z, [x12, x8, LSL #1]\n" - "inch x8\n" + "addvl x15, x15, #16\n" "fmax z31.h, p3/M, z31.h, z18.h\n" - "ld1h { z7.h }, p3/Z, [x4, #-8, MUL VL]\n" // Load from weights and bias - "cmp x8, %x[n_channels]\n" + "ld1h { z12.h }, p1/Z, [x23, x12, LSL #1]\n" + "inch x12\n" "fmax z30.h, p3/M, z30.h, z18.h\n" - "ld1h { z8.h }, p3/Z, [x4, #-7, MUL VL]\n" // Load from weights and bias - "addvl x4, x4, #-6\n" + "ld1h { z7.h }, p3/Z, [x15, #-8, MUL VL]\n" + "cmp x12, %x[n_channels]\n" "fmax z29.h, p3/M, z29.h, z18.h\n" + "ld1h { z8.h }, p3/Z, [x15, #-7, MUL VL]\n" + "addvl x15, x15, #-6\n" "fmax z28.h, p3/M, z28.h, z18.h\n" "fmin z31.h, p3/M, z31.h, z17.h\n" - "st1h { z31.h }, p0, [x22, x15, LSL #1]\n" - "mov z31.d, z16.d\n" + "st1h { z31.h }, p0, [x11, x9, LSL #1]\n" "fmin z30.h, p3/M, z30.h, z17.h\n" - "st1h { z30.h }, p0, [x21, x15, LSL #1]\n" - "mov z30.d, z16.d\n" "fmin z29.h, p3/M, z29.h, z17.h\n" - "st1h { z29.h }, p0, [x20, x15, LSL #1]\n" - "mov z29.d, z16.d\n" + "st1h { z30.h }, p0, [x10, x9, LSL #1]\n" "fmin z28.h, p3/M, z28.h, z17.h\n" - "st1h { z28.h }, p0, [x19, x15, LSL #1]\n" - "mov z28.d, z16.d\n" + "st1h { z29.h }, p0, [x28, x9, LSL #1]\n" + "st1h { z28.h }, p0, [x27, x9, LSL #1]\n" "blt 1b\n" "2:" // Channel tail - "fmla z31.h, p3/M, z4.h, z9.h\n" - "inch x15\n" - "fmla z30.h, p3/M, z3.h, z9.h\n" + "movprfx z31, z16\n fmla z31.h, p3/M, z4.h, z9.h\n" + "ldr x21, [x14, #0x28]\n" + "inch x9\n" + "movprfx z30, z16\n fmla z30.h, p3/M, z3.h, z9.h\n" + "ldr x20, [x14, #0x30]\n" "mov p0.b, p2.b\n" - "fmla z29.h, p3/M, z1.h, z9.h\n" - "fmla z28.h, p3/M, z0.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x26, x5, LSL #1]\n" + "movprfx z29, z16\n fmla z29.h, p3/M, z1.h, z9.h\n" + "ldr x19, [x14, #0x38]\n" + "movprfx z28, z16\n fmla z28.h, p3/M, z0.h, z9.h\n" + "ld1h { z9.h }, p2/Z, [x21, x13, LSL #1]\n" + "ldr x26, [x14, #0x40]\n" "fmla z31.h, p3/M, z0.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x28, x5, LSL #1]\n" + "ldr x25, [x14, #0x48]\n" "fmla z30.h, p3/M, z2.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x23, x5, LSL #1]\n" + "ld1h { z11.h }, p2/Z, [x20, x13, LSL #1]\n" "fmla z29.h, p3/M, z2.h, z12.h\n" "fmla z28.h, p3/M, z1.h, z12.h\n" + "ld1h { z10.h }, p2/Z, [x25, x13, LSL #1]\n" + "ldr x24, [x14, #0x50]\n" "fmla z31.h, p3/M, z5.h, z12.h\n" + "ldr x23, [x14, #0x58]\n" "fmla z30.h, p3/M, z4.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x7, x5, LSL #1]\n" + "ld1h { z12.h }, p2/Z, [x19, x13, LSL #1]\n" "fmla z29.h, p3/M, z6.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x17, x5, LSL #1]\n" "fmla z28.h, p3/M, z3.h, z13.h\n" + "ld1h { z9.h }, p2/Z, [x26, x13, LSL #1]\n" + "ldr x22, [x14, #0x60]\n" "fmla z31.h, p3/M, z7.h, z13.h\n" + "ldr x21, [x14, #0x68]\n" "fmla z30.h, p3/M, z6.h, z13.h\n" + "ldr x20, [x14, #0x70]\n" "fmla z29.h, p3/M, z4.h, z13.h\n" "fmla z28.h, p3/M, z8.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x14, x5, LSL #1]\n" + "ld1h { z11.h }, p2/Z, [x24, x13, LSL #1]\n" + "ldr x19, [x14, #0x78]\n" "fmla z31.h, p3/M, z1.h, z12.h\n" "fmla z30.h, p3/M, z0.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x11, x5, LSL #1]\n" + "ld1h { z12.h }, p2/Z, [x23, x13, LSL #1]\n" "fmla z29.h, p3/M, z5.h, z10.h\n" "fmla z28.h, p3/M, z4.h, z10.h\n" "fmla z31.h, p3/M, z2.h, z9.h\n" "fmla z30.h, p3/M, z1.h, z9.h\n" - "ld1h { z9.h }, p2/Z, [x10, x5, LSL #1]\n" + "ld1h { z9.h }, p2/Z, [x22, x13, LSL #1]\n" "fmla z29.h, p3/M, z0.h, z11.h\n" "fmla z28.h, p3/M, z2.h, z12.h\n" "fmla z31.h, p3/M, z8.h, z10.h\n" "fmla z30.h, p3/M, z7.h, z10.h\n" - "ld1h { z10.h }, p2/Z, [x27, x5, LSL #1]\n" + "ld1h { z10.h }, p2/Z, [x21, x13, LSL #1]\n" "fmla z29.h, p3/M, z3.h, z9.h\n" "fmla z31.h, p3/M, z3.h, z11.h\n" - "ld1h { z11.h }, p2/Z, [x25, x5, LSL #1]\n" - "fmla z30.h, p3/M, z5.h, z12.h\n" - "ld1h { z12.h }, p2/Z, [x24, x5, LSL #1]\n" + "ld1h { z11.h }, p2/Z, [x20, x13, LSL #1]\n" "fmla z28.h, p3/M, z5.h, z10.h\n" + "fmla z30.h, p3/M, z5.h, z12.h\n" + "ld1h { z12.h }, p2/Z, [x19, x13, LSL #1]\n" "fmla z29.h, p3/M, z7.h, z11.h\n" "fmla z31.h, p3/M, z6.h, z9.h\n" - "fmla z30.h, p3/M, z8.h, z10.h\n" "fmla z28.h, p3/M, z6.h, z11.h\n" + "fmla z30.h, p3/M, z8.h, z10.h\n" "fmla z29.h, p3/M, z8.h, z12.h\n" + "fmla z28.h, p3/M, z7.h, z12.h\n" "fmax z31.h, p3/M, z31.h, z18.h\n" "fmax z30.h, p3/M, z30.h, z18.h\n" - "fmla z28.h, p3/M, z7.h, z12.h\n" "fmax z29.h, p3/M, z29.h, z18.h\n" "fmin z31.h, p3/M, z31.h, z17.h\n" - "st1h { z31.h }, p0, [x22, x15, LSL #1]\n" + "st1h { z31.h }, p0, [x11, x9, LSL #1]\n" "fmin z30.h, p3/M, z30.h, z17.h\n" "fmin z29.h, p3/M, z29.h, z17.h\n" - "st1h { z30.h }, p0, [x21, x15, LSL #1]\n" + "st1h { z30.h }, p0, [x10, x9, LSL #1]\n" "fmax z28.h, p3/M, z28.h, z18.h\n" - "st1h { z29.h }, p0, [x20, x15, LSL #1]\n" + "st1h { z29.h }, p0, [x28, x9, LSL #1]\n" "fmin z28.h, p3/M, z28.h, z17.h\n" - "st1h { z28.h }, p0, [x19, x15, LSL #1]\n" + "st1h { z28.h }, p0, [x27, x9, LSL #1]\n" : : [n_channels] "r" ((unsigned long) n_channels), [offsetof_Args_inptrs] "I" (offsetof(Args, inptrs)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_args_params] "I" (offsetof(Args, params)), [params_struct] "r" (¶ms_struct) - : "cc", "memory", "p0", "p1", "p2", "p3", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z16", "z17", "z18", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z16", "z17", "z18", "z28", "z29", "z30", "z31" ); } } // 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) |