diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp | 151 |
1 files changed, 76 insertions, 75 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp index 14c07724a1..0dabc2f292 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_fp32_nhwc_max_generic_depthfirst/generic.cpp @@ -22,9 +22,10 @@ * SOFTWARE. */ -#if defined(ARM_COMPUTE_ENABLE_SME) - #include <cstdint> +#include <cstddef> + +#if defined(ARM_COMPUTE_ENABLE_SME) namespace arm_conv { namespace pooling { @@ -54,68 +55,68 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "lsr x25, %x[n_valid_cells], #0x2\n" "mov z4.s, #0xff800000\n" "mov z3.s, #0xff800000\n" - "mov x20, %x[inptrs]\n" + "mov x24, %x[inptrs]\n" "mov z2.s, #0xff800000\n" "mov z1.s, #0xff800000\n" "cbz x25, 4f\n" - "ldp x24, x23, [x20, #0x0]\n" + "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "ld1w { z0.s }, p4/Z, [x24, x9, LSL #2]\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" - "ld1w { z31.s }, p4/Z, [x23, x9, LSL #2]\n" - "ld1w { z23.s }, p4/Z, [x22, x9, LSL #2]\n" - "ld1w { z30.s }, p4/Z, [x21, x9, LSL #2]\n" - "ld1w { z18.s }, p3/Z, [x24, x28, LSL #2]\n" - "ld1w { z29.s }, p3/Z, [x23, x28, LSL #2]\n" - "ld1w { z22.s }, p3/Z, [x22, x28, LSL #2]\n" - "ld1w { z28.s }, p3/Z, [x21, x28, LSL #2]\n" - "ld1w { z17.s }, p2/Z, [x24, x27, LSL #2]\n" - "ld1w { z27.s }, p2/Z, [x23, x27, LSL #2]\n" - "ld1w { z21.s }, p2/Z, [x22, x27, LSL #2]\n" - "ld1w { z26.s }, p2/Z, [x21, x27, LSL #2]\n" - "ld1w { z16.s }, p1/Z, [x24, x26, LSL #2]\n" - "ld1w { z25.s }, p1/Z, [x23, x26, LSL #2]\n" - "ld1w { z20.s }, p1/Z, [x22, x26, LSL #2]\n" - "ld1w { z24.s }, p1/Z, [x21, x26, LSL #2]\n" + "ld1w { z0.s }, p4/Z, [x23, x9, LSL #2]\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" + "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" + "ld1w { z18.s }, p3/Z, [x23, x28, LSL #2]\n" + "ld1w { z29.s }, p3/Z, [x22, x28, LSL #2]\n" + "ld1w { z22.s }, p3/Z, [x21, x28, LSL #2]\n" + "ld1w { z28.s }, p3/Z, [x20, x28, LSL #2]\n" + "ld1w { z17.s }, p2/Z, [x23, x27, LSL #2]\n" + "ld1w { z27.s }, p2/Z, [x22, x27, LSL #2]\n" + "ld1w { z21.s }, p2/Z, [x21, x27, LSL #2]\n" + "ld1w { z26.s }, p2/Z, [x20, x27, LSL #2]\n" + "ld1w { z16.s }, p1/Z, [x23, x26, LSL #2]\n" + "ld1w { z25.s }, p1/Z, [x22, x26, LSL #2]\n" + "ld1w { z20.s }, p1/Z, [x21, x26, LSL #2]\n" + "ld1w { z24.s }, p1/Z, [x20, x26, LSL #2]\n" "beq 3f\n" "2:" // 4-vectors of channels: 4 inputs loop "movprfx z19, z0\n fmax z19.s, p0/M, z19.s, z31.s\n" "fmax z23.s, p0/M, z23.s, z30.s\n" - "ldp x24, x23, [x20, #0x0]\n" + "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" "fmax z18.s, p0/M, z18.s, z29.s\n" "fmax z22.s, p0/M, z22.s, z28.s\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" "fmax z17.s, p0/M, z17.s, z27.s\n" "fmax z21.s, p0/M, z21.s, z26.s\n" - "ld1w { z0.s }, p4/Z, [x24, x9, LSL #2]\n" + "ld1w { z0.s }, p4/Z, [x23, x9, LSL #2]\n" "fmax z16.s, p0/M, z16.s, z25.s\n" "fmax z20.s, p0/M, z20.s, z24.s\n" - "ld1w { z31.s }, p4/Z, [x23, x9, LSL #2]\n" + "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" "fmax z19.s, p0/M, z19.s, z23.s\n" "fmax z18.s, p0/M, z18.s, z22.s\n" - "ld1w { z23.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" "fmax z17.s, p0/M, z17.s, z21.s\n" "fmax z16.s, p0/M, z16.s, z20.s\n" - "ld1w { z30.s }, p4/Z, [x21, x9, LSL #2]\n" + "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" "fmax z4.s, p0/M, z4.s, z19.s\n" "fmax z3.s, p0/M, z3.s, z18.s\n" - "ld1w { z18.s }, p3/Z, [x24, x28, LSL #2]\n" + "ld1w { z18.s }, p3/Z, [x23, x28, LSL #2]\n" "fmax z2.s, p0/M, z2.s, z17.s\n" "fmax z1.s, p0/M, z1.s, z16.s\n" - "ld1w { z29.s }, p3/Z, [x23, x28, LSL #2]\n" - "ld1w { z22.s }, p3/Z, [x22, x28, LSL #2]\n" - "ld1w { z28.s }, p3/Z, [x21, x28, LSL #2]\n" - "ld1w { z17.s }, p2/Z, [x24, x27, LSL #2]\n" - "ld1w { z27.s }, p2/Z, [x23, x27, LSL #2]\n" - "ld1w { z21.s }, p2/Z, [x22, x27, LSL #2]\n" - "ld1w { z26.s }, p2/Z, [x21, x27, LSL #2]\n" - "ld1w { z16.s }, p1/Z, [x24, x26, LSL #2]\n" - "ld1w { z25.s }, p1/Z, [x23, x26, LSL #2]\n" - "ld1w { z20.s }, p1/Z, [x22, x26, LSL #2]\n" - "ld1w { z24.s }, p1/Z, [x21, x26, LSL #2]\n" + "ld1w { z29.s }, p3/Z, [x22, x28, LSL #2]\n" + "ld1w { z22.s }, p3/Z, [x21, x28, LSL #2]\n" + "ld1w { z28.s }, p3/Z, [x20, x28, LSL #2]\n" + "ld1w { z17.s }, p2/Z, [x23, x27, LSL #2]\n" + "ld1w { z27.s }, p2/Z, [x22, x27, LSL #2]\n" + "ld1w { z21.s }, p2/Z, [x21, x27, LSL #2]\n" + "ld1w { z26.s }, p2/Z, [x20, x27, LSL #2]\n" + "ld1w { z16.s }, p1/Z, [x23, x26, LSL #2]\n" + "ld1w { z25.s }, p1/Z, [x22, x26, LSL #2]\n" + "ld1w { z20.s }, p1/Z, [x21, x26, LSL #2]\n" + "ld1w { z24.s }, p1/Z, [x20, x26, LSL #2]\n" "bgt 2b\n" "3:" // 4-vectors of channels: 4 inputs tail "movprfx z19, z0\n fmax z19.s, p0/M, z19.s, z31.s\n" @@ -138,15 +139,15 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "ands x21, %x[n_valid_cells], #0x3\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x24, [x20], #0x8\n" - "ld1w { z0.s }, p4/Z, [x24, x9, LSL #2]\n" + "ldr x20, [x24], #0x8\n" + "ld1w { z16.s }, p4/Z, [x20, x9, LSL #2]\n" "subs x21, x21, #0x1\n" - "fmax z4.s, p0/M, z4.s, z0.s\n" - "ld1w { z18.s }, p3/Z, [x24, x28, LSL #2]\n" - "fmax z3.s, p0/M, z3.s, z18.s\n" - "ld1w { z17.s }, p2/Z, [x24, x27, LSL #2]\n" - "fmax z2.s, p0/M, z2.s, z17.s\n" - "ld1w { z16.s }, p1/Z, [x24, x26, LSL #2]\n" + "fmax z4.s, p0/M, z4.s, z16.s\n" + "ld1w { z16.s }, p3/Z, [x20, x28, LSL #2]\n" + "fmax z3.s, p0/M, z3.s, z16.s\n" + "ld1w { z16.s }, p2/Z, [x20, x27, LSL #2]\n" + "fmax z2.s, p0/M, z2.s, z16.s\n" + "ld1w { z16.s }, p1/Z, [x20, x26, LSL #2]\n" "fmax z1.s, p0/M, z1.s, z16.s\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End @@ -166,44 +167,44 @@ void sme_fp32_nhwc_max_generic_depthfirst_impl( "8:" // Single vector of channels: Loop "lsr x25, %x[n_valid_cells], #0x2\n" "mov z4.s, #0xff800000\n" - "mov x20, %x[inptrs]\n" + "mov x24, %x[inptrs]\n" "cbz x25, 11f\n" - "ldp x24, x23, [x20, #0x0]\n" + "ldp x20, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "ld1w { z0.s }, p4/Z, [x24, x9, LSL #2]\n" - "ldp x22, x21, [x20, #0x10]\n" - "add x20, x20, #0x20\n" - "ld1w { z31.s }, p4/Z, [x23, x9, LSL #2]\n" - "ld1w { z23.s }, p4/Z, [x22, x9, LSL #2]\n" - "ld1w { z30.s }, p4/Z, [x21, x9, LSL #2]\n" + "ld1w { z0.s }, p4/Z, [x20, x9, LSL #2]\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" + "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 4 inputs loop - "movprfx z19, z0\n fmax z19.s, p0/M, z19.s, z31.s\n" - "fmax z23.s, p0/M, z23.s, z30.s\n" - "ldp x24, x23, [x20, #0x0]\n" + "movprfx z16, z0\n fmax z16.s, p0/M, z16.s, z31.s\n" + "movprfx z17, z23\n fmax z17.s, p0/M, z17.s, z30.s\n" + "ldp x23, x22, [x24, #0x0]\n" "subs x25, x25, #0x1\n" - "fmax z19.s, p0/M, z19.s, z23.s\n" - "ldp x22, x21, [x20, #0x10]\n" - "fmax z4.s, p0/M, z4.s, z19.s\n" - "add x20, x20, #0x20\n" - "ld1w { z0.s }, p4/Z, [x24, x9, LSL #2]\n" - "ld1w { z31.s }, p4/Z, [x23, x9, LSL #2]\n" - "ld1w { z23.s }, p4/Z, [x22, x9, LSL #2]\n" - "ld1w { z30.s }, p4/Z, [x21, x9, LSL #2]\n" + "fmax z16.s, p0/M, z16.s, z17.s\n" + "ldp x21, x20, [x24, #0x10]\n" + "fmax z4.s, p0/M, z4.s, z16.s\n" + "add x24, x24, #0x20\n" + "ld1w { z0.s }, p4/Z, [x23, x9, LSL #2]\n" + "ld1w { z31.s }, p4/Z, [x22, x9, LSL #2]\n" + "ld1w { z23.s }, p4/Z, [x21, x9, LSL #2]\n" + "ld1w { z30.s }, p4/Z, [x20, x9, LSL #2]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 4 inputs tail - "movprfx z19, z0\n fmax z19.s, p0/M, z19.s, z31.s\n" - "fmax z23.s, p0/M, z23.s, z30.s\n" - "fmax z19.s, p0/M, z19.s, z23.s\n" - "fmax z4.s, p0/M, z4.s, z19.s\n" + "movprfx z16, z0\n fmax z16.s, p0/M, z16.s, z31.s\n" + "movprfx z17, z23\n fmax z17.s, p0/M, z17.s, z30.s\n" + "fmax z16.s, p0/M, z16.s, z17.s\n" + "fmax z4.s, p0/M, z4.s, z16.s\n" "11:" // Single vector of channels: Loop: After loop "ands x21, %x[n_valid_cells], #0x3\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x24, [x20], #0x8\n" - "ld1w { z0.s }, p4/Z, [x24, x9, LSL #2]\n" + "ldr x20, [x24], #0x8\n" + "ld1w { z16.s }, p4/Z, [x20, x9, LSL #2]\n" "subs x21, x21, #0x1\n" - "fmax z4.s, p0/M, z4.s, z0.s\n" + "fmax z4.s, p0/M, z4.s, z16.s\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End "st1w { z4.s }, p4, [%x[outptr], x9, LSL #2]\n" |