diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp | 514 |
1 files changed, 257 insertions, 257 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp index cda34358f5..015d0e63c2 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp @@ -25,7 +25,7 @@ #include <cstddef> #include <cstdint> -#if __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) +#if defined(ARM_COMPUTE_ENABLE_SVE) namespace arm_conv { namespace depthwise { @@ -113,7 +113,7 @@ void sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "madd x20, x8, x16, x20\n" // offset += tile_j * ld_output_col "add x9, x10, x23, LSL #2\n" "whilelt p2.s, XZR, %x[n_channels]\n" - "ld1w { z18.s }, p3/Z, [x13]\n" + "ld1w { z14.s }, p3/Z, [x13]\n" "mul x20, x20, x24\n" // offset *= output_tile_size "ld1w { z0.s }, p3/Z, [x13, #1, MUL VL]\n" "ld1w { z1.s }, p3/Z, [x13, #2, MUL VL]\n" @@ -129,10 +129,10 @@ void sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "ld1w { z6.s }, p3/Z, [x13, #7, MUL VL]\n" "addvl x13, x13, #16\n" "add x24, x11, x21, LSL #2\n" - "ld1rw { z17.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" + "ld1rw { z31.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" "cmp x15, %x[n_channels]\n" "add x23, x24, x21, LSL #2\n" - "ld1rw { z16.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" + "ld1rw { z30.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" "ld1w { z7.s }, p3/Z, [x13, #-8, MUL VL]\n" "add x22, x16, x16\n" "mov x21, #0x0\n" @@ -146,131 +146,131 @@ void sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "ld1w { z13.s }, p2/Z, [x10, x12, LSL #2]\n" "bge 3f\n" "2:" // Tile loop: Channel loop - "movprfx z24, z18\n fmla z24.s, p3/M, z7.s, z9.s\n" - "movprfx z23, z18\n fmla z23.s, p3/M, z8.s, z9.s\n" + "movprfx z29, z14\n fmla z29.s, p3/M, z7.s, z9.s\n" + "movprfx z28, z14\n fmla z28.s, p3/M, z8.s, z9.s\n" "whilelt p1.s, x15, %x[n_channels]\n" "incw x21\n" - "movprfx z25, z18\n fmla z25.s, p3/M, z6.s, z9.s\n" - "fmla z24.s, p3/M, z4.s, z13.s\n" + "movprfx z27, z14\n fmla z27.s, p3/M, z6.s, z9.s\n" + "fmla z29.s, p3/M, z4.s, z13.s\n" "incw x15\n" "mov p0.b, p2.b\n" - "movprfx z26, z18\n fmla z26.s, p3/M, z5.s, z9.s\n" - "movprfx z27, z18\n fmla z27.s, p3/M, z4.s, z9.s\n" + "movprfx z26, z14\n fmla z26.s, p3/M, z5.s, z9.s\n" + "movprfx z25, z14\n fmla z25.s, p3/M, z4.s, z9.s\n" "incw x20\n" - "movprfx z28, z18\n fmla z28.s, p3/M, z3.s, z9.s\n" - "fmla z23.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x9, x27, LSL #2]\n" - "fmla z25.s, p3/M, z2.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x9, x17, LSL #2]\n" - "movprfx z29, z18\n fmla z29.s, p3/M, z2.s, z9.s\n" - "fmla z24.s, p3/M, z6.s, z11.s\n" - "movprfx z31, z18\n fmla z31.s, p3/M, z0.s, z9.s\n" - "fmla z23.s, p3/M, z5.s, z13.s\n" - "fmla z25.s, p3/M, z3.s, z13.s\n" + "movprfx z24, z14\n fmla z24.s, p3/M, z3.s, z9.s\n" + "fmla z28.s, p3/M, z0.s, z10.s\n" + "ld1w { z23.s }, p2/Z, [x9, x27, LSL #2]\n" + "fmla z27.s, p3/M, z2.s, z11.s\n" + "ld1w { z18.s }, p2/Z, [x9, x17, LSL #2]\n" + "movprfx z22, z14\n fmla z22.s, p3/M, z2.s, z9.s\n" + "fmla z29.s, p3/M, z6.s, z18.s\n" + "movprfx z21, z14\n fmla z21.s, p3/M, z0.s, z9.s\n" + "fmla z28.s, p3/M, z5.s, z13.s\n" + "fmla z27.s, p3/M, z3.s, z13.s\n" "fmla z26.s, p3/M, z2.s, z13.s\n" - "fmla z27.s, p3/M, z1.s, z13.s\n" - "fmla z28.s, p3/M, z0.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x14, x17, LSL #2]\n" - "fmla z29.s, p3/M, z6.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x26, x25, LSL #2]\n" - "movprfx z30, z18\n fmla z30.s, p3/M, z1.s, z9.s\n" + "fmla z25.s, p3/M, z1.s, z13.s\n" "fmla z24.s, p3/M, z0.s, z13.s\n" - "ld1w { z18.s }, p3/Z, [x13]\n" - "fmla z31.s, p3/M, z8.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x14, x27, LSL #2]\n" - "fmla z23.s, p3/M, z7.s, z11.s\n" - "fmla z30.s, p3/M, z0.s, z11.s\n" - "fmla z26.s, p3/M, z4.s, z11.s\n" - "fmla z27.s, p3/M, z3.s, z11.s\n" - "fmla z29.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x10]\n" - "fmla z24.s, p3/M, z2.s, z12.s\n" - "fmla z25.s, p3/M, z1.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x28]\n" - "fmla z28.s, p3/M, z4.s, z10.s\n" - "fmla z23.s, p3/M, z1.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x10, x25, LSL #2]\n" - "fmla z30.s, p3/M, z2.s, z10.s\n" - "fmla z31.s, p3/M, z1.s, z10.s\n" - "fmla z24.s, p3/M, z8.s, z10.s\n" - "fmla z25.s, p3/M, z7.s, z10.s\n" - "fmla z27.s, p3/M, z5.s, z10.s\n" - "fmla z26.s, p3/M, z0.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x28, x12, LSL #2]\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" - "fmla z28.s, p3/M, z2.s, z13.s\n" - "fmla z30.s, p3/M, z4.s, z10.s\n" - "fmla z31.s, p3/M, z3.s, z10.s\n" - "fmla z23.s, p3/M, z3.s, z11.s\n" - "fmla z25.s, p3/M, z5.s, z13.s\n" - "ld1w { z11.s }, p2/Z, [x28, x25, LSL #2]\n" - "ld1w { z13.s }, p2/Z, [x26, x17, LSL #2]\n" - "fmla z26.s, p3/M, z6.s, z12.s\n" - "fmla z27.s, p3/M, z7.s, z10.s\n" - "ld1w { z12.s }, p2/Z, [x10, x17, LSL #2]\n" - "fmla z29.s, p3/M, z5.s, z10.s\n" - "fmla z28.s, p3/M, z6.s, z10.s\n" - "fmla z31.s, p3/M, z5.s, z11.s\n" - "fmla z30.s, p3/M, z6.s, z13.s\n" - "fmla z26.s, p3/M, z8.s, z10.s\n" - "fmla z29.s, p3/M, z7.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x26, x27, LSL #2]\n" - "fmla z24.s, p3/M, z3.s, z12.s\n" - "fmla z27.s, p3/M, z0.s, z12.s\n" - "fmla z28.s, p3/M, z8.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x10, x27, LSL #2]\n" - "fmla z30.s, p3/M, z8.s, z13.s\n" + "ld1w { z17.s }, p2/Z, [x14, x17, LSL #2]\n" + "fmla z22.s, p3/M, z6.s, z12.s\n" + "ld1w { z16.s }, p2/Z, [x26, x25, LSL #2]\n" + "movprfx z20, z14\n fmla z20.s, p3/M, z1.s, z9.s\n" + "fmla z29.s, p3/M, z0.s, z17.s\n" + "ld1w { z14.s }, p3/Z, [x13]\n" + "fmla z21.s, p3/M, z8.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x14, x27, LSL #2]\n" + "fmla z28.s, p3/M, z7.s, z18.s\n" + "fmla z20.s, p3/M, z0.s, z18.s\n" + "fmla z26.s, p3/M, z4.s, z18.s\n" + "fmla z25.s, p3/M, z3.s, z18.s\n" + "fmla z22.s, p3/M, z1.s, z18.s\n" + "ld1w { z19.s }, p2/Z, [x10]\n" + "fmla z29.s, p3/M, z2.s, z16.s\n" + "fmla z27.s, p3/M, z1.s, z16.s\n" + "ld1w { z18.s }, p2/Z, [x28]\n" + "fmla z24.s, p3/M, z4.s, z23.s\n" + "fmla z28.s, p3/M, z1.s, z17.s\n" + "ld1w { z16.s }, p2/Z, [x10, x25, LSL #2]\n" + "fmla z20.s, p3/M, z2.s, z23.s\n" + "fmla z21.s, p3/M, z1.s, z23.s\n" + "fmla z29.s, p3/M, z8.s, z23.s\n" + "fmla z27.s, p3/M, z7.s, z23.s\n" + "fmla z25.s, p3/M, z5.s, z23.s\n" + "fmla z26.s, p3/M, z0.s, z19.s\n" + "ld1w { z17.s }, p2/Z, [x28, x12, LSL #2]\n" + "fmla z22.s, p3/M, z3.s, z18.s\n" + "fmla z24.s, p3/M, z2.s, z16.s\n" + "fmla z20.s, p3/M, z4.s, z17.s\n" + "fmla z21.s, p3/M, z3.s, z17.s\n" + "fmla z28.s, p3/M, z3.s, z19.s\n" + "fmla z27.s, p3/M, z5.s, z16.s\n" + "ld1w { z19.s }, p2/Z, [x28, x25, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x26, x17, LSL #2]\n" + "fmla z26.s, p3/M, z6.s, z18.s\n" + "fmla z25.s, p3/M, z7.s, z17.s\n" + "ld1w { z18.s }, p2/Z, [x10, x17, LSL #2]\n" + "fmla z22.s, p3/M, z5.s, z17.s\n" + "fmla z24.s, p3/M, z6.s, z17.s\n" + "fmla z21.s, p3/M, z5.s, z19.s\n" + "fmla z20.s, p3/M, z6.s, z16.s\n" + "fmla z26.s, p3/M, z8.s, z17.s\n" + "fmla z22.s, p3/M, z7.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x26, x27, LSL #2]\n" + "fmla z29.s, p3/M, z3.s, z18.s\n" + "fmla z25.s, p3/M, z0.s, z18.s\n" + "fmla z24.s, p3/M, z8.s, z19.s\n" + "ld1w { z16.s }, p2/Z, [x10, x27, LSL #2]\n" + "fmla z20.s, p3/M, z8.s, z17.s\n" "addvl x10, x10, #1\n" - "fmla z31.s, p3/M, z7.s, z13.s\n" - "fmla z23.s, p3/M, z4.s, z12.s\n" - "ld1w { z13.s }, p2/Z, [x28, x27, LSL #2]\n" - "fmla z26.s, p3/M, z1.s, z12.s\n" - "fmla z24.s, p3/M, z5.s, z11.s\n" - "ld1w { z12.s }, p2/Z, [x28, x17, LSL #2]\n" + "fmla z21.s, p3/M, z7.s, z17.s\n" + "fmla z28.s, p3/M, z4.s, z18.s\n" + "ld1w { z19.s }, p2/Z, [x28, x27, LSL #2]\n" + "fmla z26.s, p3/M, z1.s, z18.s\n" + "fmla z29.s, p3/M, z5.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x28, x17, LSL #2]\n" "addvl x28, x28, #1\n" - "fmla z25.s, p3/M, z4.s, z11.s\n" - "fmla z27.s, p3/M, z2.s, z11.s\n" - "fmla z28.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x14, x12, LSL #2]\n" - "fmla z29.s, p3/M, z4.s, z12.s\n" + "fmla z27.s, p3/M, z4.s, z16.s\n" + "fmla z25.s, p3/M, z2.s, z16.s\n" + "fmla z24.s, p3/M, z1.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x14, x12, LSL #2]\n" + "fmla z22.s, p3/M, z4.s, z17.s\n" "addvl x14, x14, #1\n" - "fmla z30.s, p3/M, z3.s, z12.s\n" - "fmla z31.s, p3/M, z4.s, z13.s\n" + "fmla z20.s, p3/M, z3.s, z17.s\n" + "fmla z21.s, p3/M, z4.s, z19.s\n" "ld1w { z4.s }, p3/Z, [x13, #5, MUL VL]\n" "ld1w { z10.s }, p1/Z, [x14]\n" - "fmla z26.s, p3/M, z7.s, z12.s\n" - "fmla z27.s, p3/M, z6.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x9]\n" - "fmla z23.s, p3/M, z2.s, z11.s\n" - "fmla z24.s, p3/M, z1.s, z11.s\n" - "fmax z24.s, p3/M, z24.s, z17.s\n" + "fmla z26.s, p3/M, z7.s, z17.s\n" + "fmla z25.s, p3/M, z6.s, z17.s\n" + "ld1w { z18.s }, p2/Z, [x9]\n" + "fmla z28.s, p3/M, z2.s, z16.s\n" + "fmla z29.s, p3/M, z1.s, z16.s\n" + "fmax z29.s, p3/M, z29.s, z31.s\n" "ld1w { z1.s }, p3/Z, [x13, #2, MUL VL]\n" - "fmla z25.s, p3/M, z0.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x9, x25, LSL #2]\n" - "fmla z28.s, p3/M, z7.s, z13.s\n" + "fmla z27.s, p3/M, z0.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x9, x25, LSL #2]\n" + "fmla z24.s, p3/M, z7.s, z19.s\n" "addvl x9, x9, #1\n" - "fmla z30.s, p3/M, z5.s, z13.s\n" - "fmla z29.s, p3/M, z0.s, z12.s\n" + "fmla z20.s, p3/M, z5.s, z19.s\n" + "fmla z22.s, p3/M, z0.s, z18.s\n" "ld1w { z0.s }, p3/Z, [x13, #1, MUL VL]\n" - "fmin z24.s, p3/M, z24.s, z16.s\n" - "fmla z31.s, p3/M, z2.s, z11.s\n" - "fmla z27.s, p3/M, z8.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x26, x12, LSL #2]\n" - "fmax z27.s, p3/M, z27.s, z17.s\n" - "fmla z23.s, p3/M, z6.s, z12.s\n" - "fmla z26.s, p3/M, z3.s, z12.s\n" - "fmax z23.s, p3/M, z23.s, z17.s\n" - "fmax z26.s, p3/M, z26.s, z17.s\n" - "fmla z25.s, p3/M, z8.s, z11.s\n" - "fmla z28.s, p3/M, z5.s, z11.s\n" - "fmax z25.s, p3/M, z25.s, z17.s\n" - "fmax z28.s, p3/M, z28.s, z17.s\n" - "fmla z29.s, p3/M, z8.s, z13.s\n" - "fmla z30.s, p3/M, z7.s, z13.s\n" - "fmax z29.s, p3/M, z29.s, z17.s\n" - "fmax z30.s, p3/M, z30.s, z17.s\n" - "fmla z31.s, p3/M, z6.s, z13.s\n" - "fmax z31.s, p3/M, z31.s, z17.s\n" + "fmin z29.s, p3/M, z29.s, z30.s\n" + "fmla z21.s, p3/M, z2.s, z17.s\n" + "fmla z25.s, p3/M, z8.s, z19.s\n" + "ld1w { z16.s }, p2/Z, [x26, x12, LSL #2]\n" + "fmax z25.s, p3/M, z25.s, z31.s\n" + "fmla z28.s, p3/M, z6.s, z18.s\n" + "fmla z26.s, p3/M, z3.s, z18.s\n" + "fmax z28.s, p3/M, z28.s, z31.s\n" + "fmax z26.s, p3/M, z26.s, z31.s\n" + "fmla z27.s, p3/M, z8.s, z17.s\n" + "fmla z24.s, p3/M, z5.s, z17.s\n" + "fmax z27.s, p3/M, z27.s, z31.s\n" + "fmax z24.s, p3/M, z24.s, z31.s\n" + "fmla z22.s, p3/M, z8.s, z16.s\n" + "fmla z20.s, p3/M, z7.s, z16.s\n" + "fmax z22.s, p3/M, z22.s, z31.s\n" + "fmax z20.s, p3/M, z20.s, z31.s\n" + "fmla z21.s, p3/M, z6.s, z16.s\n" + "fmax z21.s, p3/M, z21.s, z31.s\n" "addvl x26, x26, #1\n" "ld1w { z2.s }, p3/Z, [x13, #3, MUL VL]\n" "ld1w { z3.s }, p3/Z, [x13, #4, MUL VL]\n" @@ -279,182 +279,182 @@ void sve_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "cmp x15, %x[n_channels]\n" "ld1w { z6.s }, p3/Z, [x13, #7, MUL VL]\n" "addvl x13, x13, #16\n" - "fmin z23.s, p3/M, z23.s, z16.s\n" + "fmin z28.s, p3/M, z28.s, z30.s\n" "ld1w { z9.s }, p1/Z, [x9, x12, LSL #2]\n" - "fmin z25.s, p3/M, z25.s, z16.s\n" - "fmin z26.s, p3/M, z26.s, z16.s\n" + "fmin z27.s, p3/M, z27.s, z30.s\n" + "fmin z26.s, p3/M, z26.s, z30.s\n" "ld1w { z11.s }, p1/Z, [x14, x25, LSL #2]\n" "ld1w { z12.s }, p1/Z, [x26]\n" - "fmin z27.s, p3/M, z27.s, z16.s\n" - "fmin z28.s, p3/M, z28.s, z16.s\n" + "fmin z25.s, p3/M, z25.s, z30.s\n" + "fmin z24.s, p3/M, z24.s, z30.s\n" "ld1w { z13.s }, p1/Z, [x10, x12, LSL #2]\n" - "st1w { z23.s }, p0, [x11]\n" - "fmin z29.s, p3/M, z29.s, z16.s\n" - "fmin z30.s, p3/M, z30.s, z16.s\n" - "st1w { z24.s }, p0, [x11, x16, LSL #2]\n" + "st1w { z28.s }, p0, [x11]\n" + "fmin z22.s, p3/M, z22.s, z30.s\n" + "fmin z20.s, p3/M, z20.s, z30.s\n" + "st1w { z29.s }, p0, [x11, x16, LSL #2]\n" "ld1w { z7.s }, p3/Z, [x13, #-8, MUL VL]\n" - "fmin z31.s, p3/M, z31.s, z16.s\n" - "st1w { z25.s }, p0, [x11, x22, LSL #2]\n" + "fmin z21.s, p3/M, z21.s, z30.s\n" + "st1w { z27.s }, p0, [x11, x22, LSL #2]\n" "addvl x11, x11, #1\n" "ld1w { z8.s }, p3/Z, [x13, #-7, MUL VL]\n" "st1w { z26.s }, p0, [x24]\n" "addvl x13, x13, #-6\n" - "st1w { z27.s }, p0, [x24, x16, LSL #2]\n" - "st1w { z28.s }, p0, [x24, x22, LSL #2]\n" + "st1w { z25.s }, p0, [x24, x16, LSL #2]\n" + "st1w { z24.s }, p0, [x24, x22, LSL #2]\n" "addvl x24, x24, #1\n" - "st1w { z29.s }, p0, [x23]\n" - "st1w { z30.s }, p0, [x23, x16, LSL #2]\n" - "st1w { z31.s }, p0, [x23, x22, LSL #2]\n" + "st1w { z22.s }, p0, [x23]\n" + "st1w { z20.s }, p0, [x23, x16, LSL #2]\n" + "st1w { z21.s }, p0, [x23, x22, LSL #2]\n" "addvl x23, x23, #1\n" "blt 2b\n" "3:" // Tile loop: Channel tail - "movprfx z24, z18\n fmla z24.s, p3/M, z7.s, z9.s\n" - "movprfx z23, z18\n fmla z23.s, p3/M, z8.s, z9.s\n" + "movprfx z29, z14\n fmla z29.s, p3/M, z7.s, z9.s\n" + "movprfx z28, z14\n fmla z28.s, p3/M, z8.s, z9.s\n" "ldr x8, [%x[params_struct], %[offsetof_args_tile_j]]\n" "ldr x13, [%x[params_struct], %[offsetof_args_tile_i]]\n" - "movprfx z25, z18\n fmla z25.s, p3/M, z6.s, z9.s\n" - "fmla z24.s, p3/M, z4.s, z13.s\n" + "movprfx z27, z14\n fmla z27.s, p3/M, z6.s, z9.s\n" + "fmla z29.s, p3/M, z4.s, z13.s\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" "add x8, x8, #0x1\n" - "movprfx z26, z18\n fmla z26.s, p3/M, z5.s, z9.s\n" - "movprfx z27, z18\n fmla z27.s, p3/M, z4.s, z9.s\n" + "movprfx z26, z14\n fmla z26.s, p3/M, z5.s, z9.s\n" + "movprfx z25, z14\n fmla z25.s, p3/M, z4.s, z9.s\n" "cmp x8, x20\n" "add x21, x13, #0x1\n" - "movprfx z28, z18\n fmla z28.s, p3/M, z3.s, z9.s\n" - "fmla z23.s, p3/M, z0.s, z10.s\n" - "ld1w { z10.s }, p2/Z, [x9, x27, LSL #2]\n" + "movprfx z24, z14\n fmla z24.s, p3/M, z3.s, z9.s\n" + "fmla z28.s, p3/M, z0.s, z10.s\n" + "ld1w { z23.s }, p2/Z, [x9, x27, LSL #2]\n" "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" - "fmla z25.s, p3/M, z2.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x9, x17, LSL #2]\n" - "movprfx z29, z18\n fmla z29.s, p3/M, z2.s, z9.s\n" + "fmla z27.s, p3/M, z2.s, z11.s\n" + "ld1w { z18.s }, p2/Z, [x9, x17, LSL #2]\n" + "movprfx z22, z14\n fmla z22.s, p3/M, z2.s, z9.s\n" "csel x13, x13, x21, LT\n" - "fmla z24.s, p3/M, z6.s, z11.s\n" - "movprfx z31, z18\n fmla z31.s, p3/M, z0.s, z9.s\n" + "fmla z29.s, p3/M, z6.s, z18.s\n" + "movprfx z21, z14\n fmla z21.s, p3/M, z0.s, z9.s\n" "mov p0.b, p2.b\n" "csel x8, x8, XZR, LT\n" - "fmla z23.s, p3/M, z5.s, z13.s\n" - "fmla z25.s, p3/M, z3.s, z13.s\n" + "fmla z28.s, p3/M, z5.s, z13.s\n" + "fmla z27.s, p3/M, z3.s, z13.s\n" "cmp x13, x20\n" "fmla z26.s, p3/M, z2.s, z13.s\n" - "fmla z27.s, p3/M, z1.s, z13.s\n" - "fmla z28.s, p3/M, z0.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x14, x17, LSL #2]\n" - "fmla z29.s, p3/M, z6.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x26, x25, LSL #2]\n" - "movprfx z30, z18\n fmla z30.s, p3/M, z1.s, z9.s\n" + "fmla z25.s, p3/M, z1.s, z13.s\n" "fmla z24.s, p3/M, z0.s, z13.s\n" - "fmla z31.s, p3/M, z8.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x14, x27, LSL #2]\n" - "fmla z23.s, p3/M, z7.s, z11.s\n" - "fmla z30.s, p3/M, z0.s, z11.s\n" - "fmla z26.s, p3/M, z4.s, z11.s\n" - "fmla z27.s, p3/M, z3.s, z11.s\n" - "fmla z29.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x10]\n" - "fmla z24.s, p3/M, z2.s, z12.s\n" - "fmla z25.s, p3/M, z1.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x28]\n" - "fmla z28.s, p3/M, z4.s, z10.s\n" - "fmla z23.s, p3/M, z1.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x10, x25, LSL #2]\n" - "fmla z30.s, p3/M, z2.s, z10.s\n" - "fmla z31.s, p3/M, z1.s, z10.s\n" - "fmla z24.s, p3/M, z8.s, z10.s\n" - "fmla z25.s, p3/M, z7.s, z10.s\n" - "fmla z27.s, p3/M, z5.s, z10.s\n" - "fmla z26.s, p3/M, z0.s, z11.s\n" - "ld1w { z10.s }, p2/Z, [x28, x12, LSL #2]\n" - "fmla z29.s, p3/M, z3.s, z12.s\n" - "fmla z28.s, p3/M, z2.s, z13.s\n" - "fmla z30.s, p3/M, z4.s, z10.s\n" - "fmla z31.s, p3/M, z3.s, z10.s\n" - "fmla z23.s, p3/M, z3.s, z11.s\n" - "fmla z25.s, p3/M, z5.s, z13.s\n" - "ld1w { z11.s }, p2/Z, [x28, x25, LSL #2]\n" - "ld1w { z13.s }, p2/Z, [x26, x17, LSL #2]\n" - "fmla z26.s, p3/M, z6.s, z12.s\n" - "fmla z27.s, p3/M, z7.s, z10.s\n" - "ld1w { z12.s }, p2/Z, [x10, x17, LSL #2]\n" - "fmla z29.s, p3/M, z5.s, z10.s\n" - "fmla z28.s, p3/M, z6.s, z10.s\n" - "fmla z31.s, p3/M, z5.s, z11.s\n" - "fmla z30.s, p3/M, z6.s, z13.s\n" - "fmla z26.s, p3/M, z8.s, z10.s\n" - "fmla z29.s, p3/M, z7.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x26, x27, LSL #2]\n" - "fmla z24.s, p3/M, z3.s, z12.s\n" - "fmla z27.s, p3/M, z0.s, z12.s\n" - "fmla z28.s, p3/M, z8.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x10, x27, LSL #2]\n" - "fmla z30.s, p3/M, z8.s, z13.s\n" - "fmla z31.s, p3/M, z7.s, z13.s\n" - "fmla z23.s, p3/M, z4.s, z12.s\n" - "ld1w { z13.s }, p2/Z, [x28, x27, LSL #2]\n" - "fmla z26.s, p3/M, z1.s, z12.s\n" - "fmla z24.s, p3/M, z5.s, z11.s\n" - "ld1w { z12.s }, p2/Z, [x28, x17, LSL #2]\n" - "fmla z25.s, p3/M, z4.s, z11.s\n" - "fmla z27.s, p3/M, z2.s, z11.s\n" - "fmla z28.s, p3/M, z1.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x14, x12, LSL #2]\n" - "fmla z29.s, p3/M, z4.s, z12.s\n" - "fmla z30.s, p3/M, z3.s, z12.s\n" - "fmla z31.s, p3/M, z4.s, z13.s\n" - "fmla z26.s, p3/M, z7.s, z12.s\n" - "fmla z27.s, p3/M, z6.s, z12.s\n" - "ld1w { z12.s }, p2/Z, [x9]\n" - "fmla z23.s, p3/M, z2.s, z11.s\n" - "fmla z24.s, p3/M, z1.s, z11.s\n" - "fmax z24.s, p3/M, z24.s, z17.s\n" - "fmin z24.s, p3/M, z24.s, z16.s\n" - "fmla z25.s, p3/M, z0.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x9, x25, LSL #2]\n" - "fmla z28.s, p3/M, z7.s, z13.s\n" - "fmla z30.s, p3/M, z5.s, z13.s\n" - "fmla z29.s, p3/M, z0.s, z12.s\n" - "fmla z31.s, p3/M, z2.s, z11.s\n" - "fmla z27.s, p3/M, z8.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x26, x12, LSL #2]\n" - "fmax z27.s, p3/M, z27.s, z17.s\n" - "fmla z23.s, p3/M, z6.s, z12.s\n" - "fmla z26.s, p3/M, z3.s, z12.s\n" - "fmax z23.s, p3/M, z23.s, z17.s\n" - "fmax z26.s, p3/M, z26.s, z17.s\n" - "fmla z25.s, p3/M, z8.s, z11.s\n" - "fmla z28.s, p3/M, z5.s, z11.s\n" - "fmax z25.s, p3/M, z25.s, z17.s\n" - "fmax z28.s, p3/M, z28.s, z17.s\n" - "fmla z29.s, p3/M, z8.s, z13.s\n" - "fmla z30.s, p3/M, z7.s, z13.s\n" - "fmax z29.s, p3/M, z29.s, z17.s\n" - "fmax z30.s, p3/M, z30.s, z17.s\n" - "fmla z31.s, p3/M, z6.s, z13.s\n" - "fmax z31.s, p3/M, z31.s, z17.s\n" - "fmin z23.s, p3/M, z23.s, z16.s\n" - "st1w { z23.s }, p0, [x11]\n" - "fmin z25.s, p3/M, z25.s, z16.s\n" - "fmin z26.s, p3/M, z26.s, z16.s\n" - "st1w { z24.s }, p0, [x11, x16, LSL #2]\n" - "fmin z27.s, p3/M, z27.s, z16.s\n" - "fmin z28.s, p3/M, z28.s, z16.s\n" - "st1w { z25.s }, p0, [x11, x22, LSL #2]\n" - "fmin z29.s, p3/M, z29.s, z16.s\n" - "fmin z30.s, p3/M, z30.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x14, x17, LSL #2]\n" + "fmla z22.s, p3/M, z6.s, z12.s\n" + "ld1w { z16.s }, p2/Z, [x26, x25, LSL #2]\n" + "movprfx z20, z14\n fmla z20.s, p3/M, z1.s, z9.s\n" + "fmla z29.s, p3/M, z0.s, z17.s\n" + "fmla z21.s, p3/M, z8.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x14, x27, LSL #2]\n" + "fmla z28.s, p3/M, z7.s, z18.s\n" + "fmla z20.s, p3/M, z0.s, z18.s\n" + "fmla z26.s, p3/M, z4.s, z18.s\n" + "fmla z25.s, p3/M, z3.s, z18.s\n" + "fmla z22.s, p3/M, z1.s, z18.s\n" + "ld1w { z19.s }, p2/Z, [x10]\n" + "fmla z29.s, p3/M, z2.s, z16.s\n" + "fmla z27.s, p3/M, z1.s, z16.s\n" + "ld1w { z18.s }, p2/Z, [x28]\n" + "fmla z24.s, p3/M, z4.s, z23.s\n" + "fmla z28.s, p3/M, z1.s, z17.s\n" + "ld1w { z16.s }, p2/Z, [x10, x25, LSL #2]\n" + "fmla z20.s, p3/M, z2.s, z23.s\n" + "fmla z21.s, p3/M, z1.s, z23.s\n" + "fmla z29.s, p3/M, z8.s, z23.s\n" + "fmla z27.s, p3/M, z7.s, z23.s\n" + "fmla z25.s, p3/M, z5.s, z23.s\n" + "fmla z26.s, p3/M, z0.s, z19.s\n" + "ld1w { z17.s }, p2/Z, [x28, x12, LSL #2]\n" + "fmla z22.s, p3/M, z3.s, z18.s\n" + "fmla z24.s, p3/M, z2.s, z16.s\n" + "fmla z20.s, p3/M, z4.s, z17.s\n" + "fmla z21.s, p3/M, z3.s, z17.s\n" + "fmla z28.s, p3/M, z3.s, z19.s\n" + "fmla z27.s, p3/M, z5.s, z16.s\n" + "ld1w { z19.s }, p2/Z, [x28, x25, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x26, x17, LSL #2]\n" + "fmla z26.s, p3/M, z6.s, z18.s\n" + "fmla z25.s, p3/M, z7.s, z17.s\n" + "ld1w { z18.s }, p2/Z, [x10, x17, LSL #2]\n" + "fmla z22.s, p3/M, z5.s, z17.s\n" + "fmla z24.s, p3/M, z6.s, z17.s\n" + "fmla z21.s, p3/M, z5.s, z19.s\n" + "fmla z20.s, p3/M, z6.s, z16.s\n" + "fmla z26.s, p3/M, z8.s, z17.s\n" + "fmla z22.s, p3/M, z7.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x26, x27, LSL #2]\n" + "fmla z29.s, p3/M, z3.s, z18.s\n" + "fmla z25.s, p3/M, z0.s, z18.s\n" + "fmla z24.s, p3/M, z8.s, z19.s\n" + "ld1w { z16.s }, p2/Z, [x10, x27, LSL #2]\n" + "fmla z20.s, p3/M, z8.s, z17.s\n" + "fmla z21.s, p3/M, z7.s, z17.s\n" + "fmla z28.s, p3/M, z4.s, z18.s\n" + "ld1w { z19.s }, p2/Z, [x28, x27, LSL #2]\n" + "fmla z26.s, p3/M, z1.s, z18.s\n" + "fmla z29.s, p3/M, z5.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x28, x17, LSL #2]\n" + "fmla z27.s, p3/M, z4.s, z16.s\n" + "fmla z25.s, p3/M, z2.s, z16.s\n" + "fmla z24.s, p3/M, z1.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x14, x12, LSL #2]\n" + "fmla z22.s, p3/M, z4.s, z17.s\n" + "fmla z20.s, p3/M, z3.s, z17.s\n" + "fmla z21.s, p3/M, z4.s, z19.s\n" + "fmla z26.s, p3/M, z7.s, z17.s\n" + "fmla z25.s, p3/M, z6.s, z17.s\n" + "ld1w { z18.s }, p2/Z, [x9]\n" + "fmla z28.s, p3/M, z2.s, z16.s\n" + "fmla z29.s, p3/M, z1.s, z16.s\n" + "fmax z29.s, p3/M, z29.s, z31.s\n" + "fmin z29.s, p3/M, z29.s, z30.s\n" + "fmla z27.s, p3/M, z0.s, z16.s\n" + "ld1w { z17.s }, p2/Z, [x9, x25, LSL #2]\n" + "fmla z24.s, p3/M, z7.s, z19.s\n" + "fmla z20.s, p3/M, z5.s, z19.s\n" + "fmla z22.s, p3/M, z0.s, z18.s\n" + "fmla z21.s, p3/M, z2.s, z17.s\n" + "fmla z25.s, p3/M, z8.s, z19.s\n" + "ld1w { z16.s }, p2/Z, [x26, x12, LSL #2]\n" + "fmax z25.s, p3/M, z25.s, z31.s\n" + "fmla z28.s, p3/M, z6.s, z18.s\n" + "fmla z26.s, p3/M, z3.s, z18.s\n" + "fmax z28.s, p3/M, z28.s, z31.s\n" + "fmax z26.s, p3/M, z26.s, z31.s\n" + "fmla z27.s, p3/M, z8.s, z17.s\n" + "fmla z24.s, p3/M, z5.s, z17.s\n" + "fmax z27.s, p3/M, z27.s, z31.s\n" + "fmax z24.s, p3/M, z24.s, z31.s\n" + "fmla z22.s, p3/M, z8.s, z16.s\n" + "fmla z20.s, p3/M, z7.s, z16.s\n" + "fmax z22.s, p3/M, z22.s, z31.s\n" + "fmax z20.s, p3/M, z20.s, z31.s\n" + "fmla z21.s, p3/M, z6.s, z16.s\n" + "fmax z21.s, p3/M, z21.s, z31.s\n" + "fmin z28.s, p3/M, z28.s, z30.s\n" + "st1w { z28.s }, p0, [x11]\n" + "fmin z27.s, p3/M, z27.s, z30.s\n" + "fmin z26.s, p3/M, z26.s, z30.s\n" + "st1w { z29.s }, p0, [x11, x16, LSL #2]\n" + "fmin z25.s, p3/M, z25.s, z30.s\n" + "fmin z24.s, p3/M, z24.s, z30.s\n" + "st1w { z27.s }, p0, [x11, x22, LSL #2]\n" + "fmin z22.s, p3/M, z22.s, z30.s\n" + "fmin z20.s, p3/M, z20.s, z30.s\n" "st1w { z26.s }, p0, [x24]\n" - "fmin z31.s, p3/M, z31.s, z16.s\n" - "st1w { z27.s }, p0, [x24, x16, LSL #2]\n" - "st1w { z28.s }, p0, [x24, x22, LSL #2]\n" - "st1w { z29.s }, p0, [x23]\n" - "st1w { z30.s }, p0, [x23, x16, LSL #2]\n" - "st1w { z31.s }, p0, [x23, x22, LSL #2]\n" + "fmin z21.s, p3/M, z21.s, z30.s\n" + "st1w { z25.s }, p0, [x24, x16, LSL #2]\n" + "st1w { z24.s }, p0, [x24, x22, LSL #2]\n" + "st1w { z22.s }, p0, [x23]\n" + "st1w { z20.s }, p0, [x23, x16, LSL #2]\n" + "st1w { z21.s }, p0, [x23, x22, LSL #2]\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) - : "cc", "memory", "p0", "p1", "p2", "p3", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "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", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } } // namespace depthwise } // namespace arm_conv -#endif // __aarch64__ && defined(ARM_COMPUTE_ENABLE_SVE) +#endif // defined(ARM_COMPUTE_ENABLE_SVE) |