diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp | 1890 |
1 files changed, 972 insertions, 918 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp index eb684e2118..67a6eb32bb 100644 --- a/src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp +++ b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp @@ -26,11 +26,20 @@ #ifdef __ARM_FEATURE_SVE template<> -inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const uint32_t alpha, const uint32_t beta) +void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const uint32_t *bias, Activation act, bool append) { + UNUSED(act); + const uint32_t *inptr = in; + uint32_t nullbias[192] = { 0 }; + + if (!append && !bias) + { + memset(nullbias, 0, (3 * get_vector_length<uint32_t>() * sizeof(uint32_t))); + } - for (int y=y0; y<ymax; y+=8) { + for (int y=y0; y<ymax; y+=8) + { uint32_t *outptr0 = out + (y * ldout) + x0; uint32_t *outptr1 = outptr0 + ldout; uint32_t *outptr2 = outptr1 + ldout; @@ -42,42 +51,44 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in const int height = ymax - y; - for (int i=x0; i<xmax; i+=(3 * get_vector_length<uint32_t>())) { - if (beta==0u) + for (int i=x0; i<xmax; i+=(3 * get_vector_length<uint32_t>())) + { + if (append) { - switch(height) { + switch(height) + { case 1: { long w = xmax - i; long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" - "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" - "addvl %[outptr0], %[outptr0], #3\n" - "1:\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" "addvl %[inptr], %[inptr], #24\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -88,41 +99,47 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" - "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z13.s, p0/z, [%[inptr], #3, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z6.s, p1/z, [%[outptr1], #1, MUL VL]\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "ld1w z14.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z7.s, p2/z, [%[outptr1], #2, MUL VL]\n" + "ld1w z15.s, p2/z, [%[inptr], #5, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" "addvl %[outptr0], %[outptr0], #3\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p0, [%[outptr1]]\n" + "st1w z14.s, p1, [%[outptr1], #1, MUL VL]\n" + "st1w z15.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -133,50 +150,62 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" - "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z13.s, p0/z, [%[inptr], #3, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z6.s, p1/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z14.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z7.s, p2/z, [%[outptr1], #2, MUL VL]\n" + "ld1w z15.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z8.s, p0/z, [%[outptr2]]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #6, MUL VL]\n" "addvl %[outptr0], %[outptr0], #3\n" - "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" - "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p1/z, [%[outptr2], #1, MUL VL]\n" + "add z16.s, z16.s, z8.s\n" + "ld1w z17.s, p1/z, [%[inptr], #7, MUL VL]\n" + "ld1w z2.s, p2/z, [%[outptr2], #2, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "st1w z14.s, p1, [%[outptr1], #1, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "ld1w z10.s, p2/z, [x8, #-8, MUL VL]\n" + "st1w z15.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "st1w z16.s, p0, [%[outptr2]]\n" + "st1w z17.s, p1, [%[outptr2], #1, MUL VL]\n" + "st1w z10.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -187,58 +216,76 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z13.s, p0/z, [%[inptr], #3, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z6.s, p1/z, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z14.s, p1/z, [%[inptr], #4, MUL VL]\n" + "ld1w z7.s, p2/z, [%[outptr1], #2, MUL VL]\n" + "ld1w z15.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z8.s, p0/z, [%[outptr2]]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #6, MUL VL]\n" "addvl %[outptr0], %[outptr0], #3\n" - "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p1/z, [%[outptr2], #1, MUL VL]\n" + "add z16.s, z16.s, z8.s\n" + "ld1w z17.s, p1/z, [%[inptr], #7, MUL VL]\n" + "ld1w z2.s, p2/z, [%[outptr2], #2, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "st1w z14.s, p1, [%[outptr1], #1, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "ld1w z10.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z3.s, p0/z, [%[outptr3]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "st1w z15.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z4.s, p1/z, [%[outptr3], #1, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z16.s, p0, [%[outptr2]]\n" + "ld1w z12.s, p1/z, [x8, #-6, MUL VL]\n" + "ld1w z5.s, p2/z, [%[outptr3], #2, MUL VL]\n" + "ld1w z13.s, p2/z, [x8, #-5, MUL VL]\n" + "st1w z17.s, p1, [%[outptr2], #1, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "add z13.s, z13.s, z5.s\n" + "st1w z10.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z12.s, p1, [%[outptr3], #1, MUL VL]\n" + "st1w z13.s, p2, [%[outptr3], #2, MUL VL]\n" "addvl %[outptr3], %[outptr3], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -249,67 +296,91 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z11.s, p0/z, [%[inptr], #7, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-6, MUL VL]\n" - "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #-3, MUL VL]\n" - "st1w z10.s, p0, [%[outptr1], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr2], #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr3], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr4], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z13.s, p0/z, [%[inptr], #3, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z6.s, p1/z, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z14.s, p1/z, [%[inptr], #4, MUL VL]\n" + "ld1w z7.s, p2/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z15.s, p2/z, [%[inptr], #5, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [%[outptr2]]\n" "addvl %[outptr0], %[outptr0], #3\n" - "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" - "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p0, [%[outptr1]]\n" + "ld1w z16.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z9.s, p1/z, [%[outptr2], #1, MUL VL]\n" + "ld1w z17.s, p1/z, [%[inptr], #7, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p1, [%[outptr1], #1, MUL VL]\n" + "ld1w z2.s, p2/z, [%[outptr2], #2, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "ld1w z10.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z3.s, p0/z, [%[outptr3]]\n" + "st1w z15.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" - "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p1/z, [%[outptr3], #1, MUL VL]\n" + "st1w z16.s, p0, [%[outptr2]]\n" + "ld1w z12.s, p1/z, [x8, #-6, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "ld1w z5.s, p2/z, [%[outptr3], #2, MUL VL]\n" + "ld1w z13.s, p2/z, [x8, #-5, MUL VL]\n" + "st1w z17.s, p1, [%[outptr2], #1, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z6.s, p0/z, [%[outptr4]]\n" + "ld1w z14.s, p0/z, [x8, #-4, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z10.s, p2, [%[outptr2], #2, MUL VL]\n" + "ld1w z7.s, p1/z, [%[outptr4], #1, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" - "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "ld1w z15.s, p1/z, [x8, #-3, MUL VL]\n" + "ld1w z8.s, p2/z, [%[outptr4], #2, MUL VL]\n" + "ld1w z16.s, p2/z, [x8, #-2, MUL VL]\n" + "st1w z12.s, p1, [%[outptr3], #1, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "add z16.s, z16.s, z8.s\n" + "st1w z13.s, p2, [%[outptr3], #2, MUL VL]\n" "addvl %[outptr3], %[outptr3], #3\n" - "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "st1w z14.s, p0, [%[outptr4]]\n" + "st1w z15.s, p1, [%[outptr4], #1, MUL VL]\n" + "st1w z16.s, p2, [%[outptr4], #2, MUL VL]\n" "addvl %[outptr4], %[outptr4], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -320,76 +391,106 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "st1w z9.s, p0, [%[outptr5]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x280]\n" - "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z8.s, p0/z, [%[inptr], #7, MUL VL]\n" - "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #-6, MUL VL]\n" - "ld1w z10.s, p0/z, [x8, #-3, MUL VL]\n" - "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [x8]\n" - "st1w z8.s, p0, [%[outptr2], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr3], #1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr4], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr5], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z13.s, p0/z, [%[inptr], #3, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z6.s, p1/z, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z14.s, p1/z, [%[inptr], #4, MUL VL]\n" + "ld1w z7.s, p2/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z15.s, p2/z, [%[inptr], #5, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [%[outptr2]]\n" "addvl %[outptr0], %[outptr0], #3\n" - "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p0, [%[outptr1]]\n" + "ld1w z16.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z9.s, p1/z, [%[outptr2], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p1, [%[outptr1], #1, MUL VL]\n" + "ld1w z17.s, p1/z, [%[inptr], #7, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "ld1w z2.s, p2/z, [%[outptr2], #2, MUL VL]\n" + "st1w z15.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" - "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "ld1w z10.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z3.s, p0/z, [%[outptr3]]\n" + "st1w z16.s, p0, [%[outptr2]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z4.s, p1/z, [%[outptr3], #1, MUL VL]\n" + "ld1w z12.s, p1/z, [x8, #-6, MUL VL]\n" + "st1w z17.s, p1, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "ld1w z5.s, p2/z, [%[outptr3], #2, MUL VL]\n" + "ld1w z13.s, p2/z, [x8, #-5, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "st1w z10.s, p2, [%[outptr2], #2, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr4]]\n" "addvl %[outptr2], %[outptr2], #3\n" - "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" - "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "ld1w z14.s, p0/z, [x8, #-4, MUL VL]\n" + "ld1w z7.s, p1/z, [%[outptr4], #1, MUL VL]\n" + "ld1w z15.s, p1/z, [x8, #-3, MUL VL]\n" + "st1w z12.s, p1, [%[outptr3], #1, MUL VL]\n" + "add z14.s, z14.s, z6.s\n" + "ld1w z8.s, p2/z, [%[outptr4], #2, MUL VL]\n" + "ld1w z16.s, p2/z, [x8, #-2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p2, [%[outptr3], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [%[outptr5]]\n" "addvl %[outptr3], %[outptr3], #3\n" - "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p0, [%[outptr4]]\n" + "ld1w z17.s, p0/z, [x8, #-1, MUL VL]\n" + "ld1w z2.s, p1/z, [%[outptr5], #1, MUL VL]\n" + "ld1w z10.s, p1/z, [x8]\n" + "st1w z15.s, p1, [%[outptr4], #1, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "ld1w z3.s, p2/z, [%[outptr5], #2, MUL VL]\n" + "ld1w z11.s, p2/z, [x8, #1, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "st1w z16.s, p2, [%[outptr4], #2, MUL VL]\n" "addvl %[outptr4], %[outptr4], #3\n" - "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z17.s, p0, [%[outptr5]]\n" + "st1w z10.s, p1, [%[outptr5], #1, MUL VL]\n" + "st1w z11.s, p2, [%[outptr5], #2, MUL VL]\n" "addvl %[outptr5], %[outptr5], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -400,85 +501,121 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "st1w z9.s, p0, [%[outptr5]]\n" - "st1w z10.s, p0, [%[outptr6]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x280]\n" - "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" - "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z10.s, p0/z, [x8, #-6, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-3, MUL VL]\n" - "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" - "ld1w z8.s, p0/z, [x8]\n" - "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #3, MUL VL]\n" - "st1w z10.s, p0, [%[outptr3], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr4], #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr5], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr6], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z13.s, p0/z, [%[inptr], #3, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z6.s, p1/z, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z14.s, p1/z, [%[inptr], #4, MUL VL]\n" + "ld1w z7.s, p2/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z15.s, p2/z, [%[inptr], #5, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [%[outptr2]]\n" "addvl %[outptr0], %[outptr0], #3\n" - "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" - "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p0, [%[outptr1]]\n" + "ld1w z16.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z9.s, p1/z, [%[outptr2], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p1, [%[outptr1], #1, MUL VL]\n" + "ld1w z17.s, p1/z, [%[inptr], #7, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "ld1w z2.s, p2/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add z17.s, z17.s, z9.s\n" + "st1w z15.s, p2, [%[outptr1], #2, MUL VL]\n" + "ld1w z10.s, p2/z, [x8, #-8, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" - "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "ld1w z3.s, p0/z, [%[outptr3]]\n" + "addvl %[inptr], %[inptr], #24\n" + "add z10.s, z10.s, z2.s\n" + "st1w z16.s, p0, [%[outptr2]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p1/z, [%[outptr3], #1, MUL VL]\n" + "ld1w z12.s, p1/z, [x8, #-6, MUL VL]\n" + "st1w z17.s, p1, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "ld1w z5.s, p2/z, [%[outptr3], #2, MUL VL]\n" + "ld1w z13.s, p2/z, [x8, #-5, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "st1w z10.s, p2, [%[outptr2], #2, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr4]]\n" "addvl %[outptr2], %[outptr2], #3\n" - "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" - "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "ld1w z14.s, p0/z, [x8, #-4, MUL VL]\n" + "ld1w z7.s, p1/z, [%[outptr4], #1, MUL VL]\n" + "ld1w z15.s, p1/z, [x8, #-3, MUL VL]\n" + "st1w z12.s, p1, [%[outptr3], #1, MUL VL]\n" + "add z14.s, z14.s, z6.s\n" + "ld1w z8.s, p2/z, [%[outptr4], #2, MUL VL]\n" + "ld1w z16.s, p2/z, [x8, #-2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p2, [%[outptr3], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [%[outptr5]]\n" "addvl %[outptr3], %[outptr3], #3\n" - "ld1w z11.s, p0/z, [x8, #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" - "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p0, [%[outptr4]]\n" + "ld1w z17.s, p0/z, [x8, #-1, MUL VL]\n" + "ld1w z2.s, p1/z, [%[outptr5], #1, MUL VL]\n" + "ld1w z10.s, p1/z, [x8]\n" + "st1w z15.s, p1, [%[outptr4], #1, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "ld1w z3.s, p2/z, [%[outptr5], #2, MUL VL]\n" + "ld1w z11.s, p2/z, [x8, #1, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "st1w z16.s, p2, [%[outptr4], #2, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr6]]\n" "addvl %[outptr4], %[outptr4], #3\n" - "ld1w z8.s, p0/z, [x8, #4, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" - "st1w z11.s, p0, [%[outptr5], #2, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z17.s, p0, [%[outptr5]]\n" + "ld1w z12.s, p0/z, [x8, #2, MUL VL]\n" + "ld1w z5.s, p1/z, [%[outptr6], #1, MUL VL]\n" + "ld1w z13.s, p1/z, [x8, #3, MUL VL]\n" + "st1w z10.s, p1, [%[outptr5], #1, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z6.s, p2/z, [%[outptr6], #2, MUL VL]\n" + "ld1w z14.s, p2/z, [x8, #4, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p2, [%[outptr5], #2, MUL VL]\n" "addvl %[outptr5], %[outptr5], #3\n" - "st1w z8.s, p0, [%[outptr6], #2, MUL VL]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p0, [%[outptr6]]\n" + "st1w z13.s, p1, [%[outptr6], #1, MUL VL]\n" + "st1w z14.s, p2, [%[outptr6], #2, MUL VL]\n" "addvl %[outptr6], %[outptr6], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -490,93 +627,135 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "ld1w z11.s, p0/z, [x8, #5, MUL VL]\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "st1w z9.s, p0, [%[outptr5]]\n" - "st1w z10.s, p0, [%[outptr6]]\n" - "st1w z11.s, p0, [%[outptr7]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[outptr0]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z10.s, p0/z, [%[inptr]]\n" "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x280]\n" - "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" - "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-3, MUL VL]\n" - "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [x8]\n" - "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" - "ld1w z10.s, p0/z, [x8, #3, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr4], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr5], #1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr6], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr7], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" - "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "ld1w z3.s, p1/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z13.s, p0/z, [%[inptr], #3, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z10.s, p0, [%[outptr0]]\n" + "ld1w z4.s, p2/z, [%[outptr0], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p1, [%[outptr0], #1, MUL VL]\n" + "ld1w z12.s, p2/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z6.s, p1/z, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z14.s, p1/z, [%[inptr], #4, MUL VL]\n" + "ld1w z7.s, p2/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z15.s, p2/z, [%[inptr], #5, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p2, [%[outptr0], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [%[outptr2]]\n" "addvl %[outptr0], %[outptr0], #3\n" - "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p0, [%[outptr1]]\n" + "ld1w z16.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z9.s, p1/z, [%[outptr2], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p1, [%[outptr1], #1, MUL VL]\n" + "ld1w z17.s, p1/z, [%[inptr], #7, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "ld1w z2.s, p2/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add z17.s, z17.s, z9.s\n" + "st1w z15.s, p2, [%[outptr1], #2, MUL VL]\n" + "ld1w z10.s, p2/z, [x8, #-8, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" - "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "ld1w z3.s, p0/z, [%[outptr3]]\n" + "prfm PLDL1KEEP, [%[outptr7], #0x60]\n" + "add z10.s, z10.s, z2.s\n" + "st1w z16.s, p0, [%[outptr2]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "ld1w z4.s, p1/z, [%[outptr3], #1, MUL VL]\n" + "st1w z17.s, p1, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "ld1w z12.s, p1/z, [x8, #-6, MUL VL]\n" + "ld1w z5.s, p2/z, [%[outptr3], #2, MUL VL]\n" + "ld1w z13.s, p2/z, [x8, #-5, MUL VL]\n" + "st1w z10.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" - "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z6.s, p0/z, [%[outptr4]]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "ld1w z14.s, p0/z, [x8, #-4, MUL VL]\n" + "ld1w z7.s, p1/z, [%[outptr4], #1, MUL VL]\n" + "ld1w z15.s, p1/z, [x8, #-3, MUL VL]\n" + "st1w z12.s, p1, [%[outptr3], #1, MUL VL]\n" + "add z14.s, z14.s, z6.s\n" + "ld1w z8.s, p2/z, [%[outptr4], #2, MUL VL]\n" + "ld1w z16.s, p2/z, [x8, #-2, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "st1w z13.s, p2, [%[outptr3], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [%[outptr5]]\n" "addvl %[outptr3], %[outptr3], #3\n" - "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" - "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p0, [%[outptr4]]\n" + "ld1w z17.s, p0/z, [x8, #-1, MUL VL]\n" + "ld1w z2.s, p1/z, [%[outptr5], #1, MUL VL]\n" + "ld1w z10.s, p1/z, [x8]\n" + "st1w z15.s, p1, [%[outptr4], #1, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "ld1w z3.s, p2/z, [%[outptr5], #2, MUL VL]\n" + "ld1w z11.s, p2/z, [x8, #1, MUL VL]\n" + "add z10.s, z10.s, z2.s\n" + "st1w z16.s, p2, [%[outptr4], #2, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr6]]\n" "addvl %[outptr4], %[outptr4], #3\n" - "ld1w z10.s, p0/z, [x8, #4, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" - "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "add z11.s, z11.s, z3.s\n" + "st1w z17.s, p0, [%[outptr5]]\n" + "ld1w z12.s, p0/z, [x8, #2, MUL VL]\n" + "ld1w z5.s, p1/z, [%[outptr6], #1, MUL VL]\n" + "ld1w z13.s, p1/z, [x8, #3, MUL VL]\n" + "st1w z10.s, p1, [%[outptr5], #1, MUL VL]\n" + "add z12.s, z12.s, z4.s\n" + "ld1w z6.s, p2/z, [%[outptr6], #2, MUL VL]\n" + "ld1w z14.s, p2/z, [x8, #4, MUL VL]\n" + "add z13.s, z13.s, z5.s\n" + "st1w z11.s, p2, [%[outptr5], #2, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr7]]\n" "addvl %[outptr5], %[outptr5], #3\n" - "ld1w z11.s, p0/z, [x8, #7, MUL VL]\n" - "prfm PSTL1KEEP, [%[outptr7], #0x60]\n" - "st1w z10.s, p0, [%[outptr6], #2, MUL VL]\n" + "add z14.s, z14.s, z6.s\n" + "st1w z12.s, p0, [%[outptr6]]\n" + "ld1w z15.s, p0/z, [x8, #5, MUL VL]\n" + "ld1w z8.s, p1/z, [%[outptr7], #1, MUL VL]\n" + "ld1w z16.s, p1/z, [x8, #6, MUL VL]\n" + "st1w z13.s, p1, [%[outptr6], #1, MUL VL]\n" + "add z15.s, z15.s, z7.s\n" + "ld1w z9.s, p2/z, [%[outptr7], #2, MUL VL]\n" + "ld1w z17.s, p2/z, [x8, #7, MUL VL]\n" + "add z16.s, z16.s, z8.s\n" + "st1w z14.s, p2, [%[outptr6], #2, MUL VL]\n" "addvl %[outptr6], %[outptr6], #3\n" - "st1w z11.s, p0, [%[outptr7], #2, MUL VL]\n" + "add z17.s, z17.s, z9.s\n" + "st1w z15.s, p0, [%[outptr7]]\n" + "st1w z16.s, p1, [%[outptr7], #1, MUL VL]\n" + "st1w z17.s, p2, [%[outptr7], #2, MUL VL]\n" "addvl %[outptr7], %[outptr7], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -586,45 +765,46 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in } else { - switch(height) { + const uint32_t *biasptr = nullbias; + if (bias) + { + biasptr = bias + i; + } + + switch(height) + { case 1: { long w = xmax - i; long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "add z8.s, z8.s, z4.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" - "ld1w z5.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" - "addvl %[outptr0], %[outptr0], #3\n" - "1:\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "whilelt p2.s, %[p], %[w]\n" + "add z13.s, z13.s, z2.s\n" + "add z14.s, z14.s, z3.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" "addvl %[inptr], %[inptr], #24\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z15.s, z15.s, z4.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -635,53 +815,44 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z5.s, p0/z, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z6.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr1], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "add z11.s, z11.s, z7.s\n" - "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "add z13.s, z13.s, z2.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #3, MUL VL]\n" + "add z14.s, z14.s, z3.s\n" + "ld1w z17.s, p1/z, [%[inptr], #4, MUL VL]\n" + "ld1w z18.s, p2/z, [%[inptr], #5, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "add z15.s, z15.s, z4.s\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z16.s, z16.s, z2.s\n" + "add z17.s, z17.s, z3.s\n" + "add z18.s, z18.s, z4.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" "addvl %[outptr0], %[outptr0], #3\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "st1w z16.s, p0, [%[outptr1]]\n" + "st1w z17.s, p1, [%[outptr1], #1, MUL VL]\n" + "st1w z18.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -692,68 +863,56 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z5.s, p0/z, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z7.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr1], #1, MUL VL]\n" - "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr2], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "add z9.s, z9.s, z5.s\n" - "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "add z13.s, z13.s, z2.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" - "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" - "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "add z14.s, z14.s, z3.s\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z15.s, z15.s, z4.s\n" + "ld1w z17.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z16.s, z16.s, z2.s\n" + "ld1w z18.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z19.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "add z17.s, z17.s, z3.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "add z18.s, z18.s, z4.s\n" + "ld1w z20.s, p1/z, [%[inptr], #7, MUL VL]\n" + "add z19.s, z19.s, z2.s\n" + "ld1w z13.s, p2/z, [x8, #-8, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" "addvl %[outptr0], %[outptr0], #3\n" - "add z8.s, z8.s, z4.s\n" - "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "add z20.s, z20.s, z3.s\n" + "add z13.s, z13.s, z4.s\n" + "st1w z16.s, p0, [%[outptr1]]\n" + "st1w z17.s, p1, [%[outptr1], #1, MUL VL]\n" + "st1w z18.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "st1w z19.s, p0, [%[outptr2]]\n" + "st1w z20.s, p1, [%[outptr2], #1, MUL VL]\n" + "st1w z13.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -764,82 +923,67 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z5.s, p0/z, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3]]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z4.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr1], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2], #1, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "add z13.s, z13.s, z2.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" - "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" - "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "add z14.s, z14.s, z3.s\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z15.s, z15.s, z4.s\n" + "ld1w z17.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z16.s, z16.s, z2.s\n" + "ld1w z18.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z19.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "add z17.s, z17.s, z3.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "add z18.s, z18.s, z4.s\n" + "ld1w z20.s, p1/z, [%[inptr], #7, MUL VL]\n" + "add z19.s, z19.s, z2.s\n" + "ld1w z13.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z14.s, p0/z, [x8, #-7, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "add z20.s, z20.s, z3.s\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z4.s\n" + "ld1w z15.s, p1/z, [x8, #-6, MUL VL]\n" + "add z14.s, z14.s, z2.s\n" "addvl %[outptr0], %[outptr0], #3\n" - "add z10.s, z10.s, z6.s\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "st1w z16.s, p0, [%[outptr1]]\n" + "addvl %[inptr], %[inptr], #24\n" + "add z15.s, z15.s, z3.s\n" + "ld1w z16.s, p2/z, [x8, #-5, MUL VL]\n" + "st1w z17.s, p1, [%[outptr1], #1, MUL VL]\n" + "add z16.s, z16.s, z4.s\n" + "st1w z18.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "add z11.s, z11.s, z7.s\n" - "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "st1w z19.s, p0, [%[outptr2]]\n" + "st1w z20.s, p1, [%[outptr2], #1, MUL VL]\n" + "st1w z13.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "st1w z14.s, p0, [%[outptr3]]\n" + "st1w z15.s, p1, [%[outptr3], #1, MUL VL]\n" + "st1w z16.s, p2, [%[outptr3], #2, MUL VL]\n" "addvl %[outptr3], %[outptr3], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -850,97 +994,79 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z5.s, p0/z, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3]]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr4]]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "add z8.s, z8.s, z4.s\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z5.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr1], #1, MUL VL]\n" - "ld1w z10.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr2], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [%[inptr], #7, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr3], #1, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z8.s, p0/z, [x8, #-6, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr4], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #-3, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "st1w z10.s, p0, [%[outptr1], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z11.s, p0, [%[outptr2], #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr3], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr4], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "add z13.s, z13.s, z2.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" - "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" - "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "add z14.s, z14.s, z3.s\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z15.s, z15.s, z4.s\n" + "ld1w z17.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z16.s, z16.s, z2.s\n" + "ld1w z18.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z19.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "add z17.s, z17.s, z3.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "add z18.s, z18.s, z4.s\n" + "ld1w z20.s, p1/z, [%[inptr], #7, MUL VL]\n" + "add z19.s, z19.s, z2.s\n" + "ld1w z13.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z14.s, p0/z, [x8, #-7, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "add z20.s, z20.s, z3.s\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z4.s\n" + "ld1w z15.s, p1/z, [x8, #-6, MUL VL]\n" + "add z14.s, z14.s, z2.s\n" "addvl %[outptr0], %[outptr0], #3\n" - "add z8.s, z8.s, z4.s\n" - "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "st1w z16.s, p0, [%[outptr1]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "add z15.s, z15.s, z3.s\n" + "ld1w z16.s, p2/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z17.s, p1, [%[outptr1], #1, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "add z16.s, z16.s, z4.s\n" + "ld1w z17.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z18.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z6.s, p0/z, [%[outptr4], #2, MUL VL]\n" - "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "add z17.s, z17.s, z2.s\n" + "ld1w z18.s, p1/z, [x8, #-3, MUL VL]\n" + "st1w z19.s, p0, [%[outptr2]]\n" + "ld1w z19.s, p2/z, [x8, #-2, MUL VL]\n" + "add z18.s, z18.s, z3.s\n" + "st1w z20.s, p1, [%[outptr2], #1, MUL VL]\n" + "add z19.s, z19.s, z4.s\n" + "st1w z13.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "st1w z14.s, p0, [%[outptr3]]\n" + "st1w z15.s, p1, [%[outptr3], #1, MUL VL]\n" + "st1w z16.s, p2, [%[outptr3], #2, MUL VL]\n" "addvl %[outptr3], %[outptr3], #3\n" - "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "st1w z17.s, p0, [%[outptr4]]\n" + "st1w z18.s, p1, [%[outptr4], #1, MUL VL]\n" + "st1w z19.s, p2, [%[outptr4], #2, MUL VL]\n" "addvl %[outptr4], %[outptr4], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -951,112 +1077,91 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z5.s, p0/z, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3]]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr4]]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z5.s, p0/z, [%[outptr5]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "st1w z9.s, p0, [%[outptr5]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z6.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x280]\n" - "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr1], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr2], #1, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z8.s, p0/z, [%[inptr], #7, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr3], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z9.s, p0/z, [x8, #-6, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr4], #1, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z10.s, p0/z, [x8, #-3, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z7.s, p0/z, [%[outptr5], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z11.s, p0/z, [x8]\n" - "st1w z8.s, p0, [%[outptr2], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z9.s, p0, [%[outptr3], #1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr4], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr5], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "add z13.s, z13.s, z2.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" - "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" - "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "add z14.s, z14.s, z3.s\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z15.s, z15.s, z4.s\n" + "ld1w z17.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z16.s, z16.s, z2.s\n" + "ld1w z18.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z19.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "add z17.s, z17.s, z3.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "add z18.s, z18.s, z4.s\n" + "ld1w z20.s, p1/z, [%[inptr], #7, MUL VL]\n" + "add z19.s, z19.s, z2.s\n" + "ld1w z13.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z14.s, p0/z, [x8, #-7, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "add z20.s, z20.s, z3.s\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z4.s\n" + "ld1w z15.s, p1/z, [x8, #-6, MUL VL]\n" + "add z14.s, z14.s, z2.s\n" "addvl %[outptr0], %[outptr0], #3\n" - "add z10.s, z10.s, z6.s\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "st1w z16.s, p0, [%[outptr1]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "add z15.s, z15.s, z3.s\n" + "ld1w z16.s, p2/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z17.s, p1, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "add z16.s, z16.s, z4.s\n" + "ld1w z17.s, p0/z, [x8, #-4, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z18.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z4.s, p0/z, [%[outptr4], #2, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "add z17.s, z17.s, z2.s\n" + "ld1w z18.s, p1/z, [x8, #-3, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "st1w z19.s, p0, [%[outptr2]]\n" + "ld1w z19.s, p2/z, [x8, #-2, MUL VL]\n" + "add z18.s, z18.s, z3.s\n" + "st1w z20.s, p1, [%[outptr2], #1, MUL VL]\n" + "add z19.s, z19.s, z4.s\n" + "ld1w z20.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z13.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr5], #2, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "add z20.s, z20.s, z2.s\n" + "ld1w z13.s, p1/z, [x8]\n" + "st1w z14.s, p0, [%[outptr3]]\n" + "ld1w z14.s, p2/z, [x8, #1, MUL VL]\n" + "add z13.s, z13.s, z3.s\n" + "st1w z15.s, p1, [%[outptr3], #1, MUL VL]\n" + "add z14.s, z14.s, z4.s\n" + "st1w z16.s, p2, [%[outptr3], #2, MUL VL]\n" "addvl %[outptr3], %[outptr3], #3\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "st1w z17.s, p0, [%[outptr4]]\n" + "st1w z18.s, p1, [%[outptr4], #1, MUL VL]\n" + "st1w z19.s, p2, [%[outptr4], #2, MUL VL]\n" "addvl %[outptr4], %[outptr4], #3\n" - "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "st1w z20.s, p0, [%[outptr5]]\n" + "st1w z13.s, p1, [%[outptr5], #1, MUL VL]\n" + "st1w z14.s, p2, [%[outptr5], #2, MUL VL]\n" "addvl %[outptr5], %[outptr5], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -1067,127 +1172,103 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z5.s, p0/z, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3]]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr4]]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z5.s, p0/z, [%[outptr5]]\n" - "ld1w z6.s, p0/z, [%[outptr6]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "st1w z9.s, p0, [%[outptr5]]\n" - "st1w z10.s, p0, [%[outptr6]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z7.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x280]\n" - "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr1], #1, MUL VL]\n" - "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr2], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr3], #1, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [x8, #-6, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr4], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-3, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z4.s, p0/z, [%[outptr5], #1, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr6], #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z8.s, p0/z, [x8]\n" - "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [x8, #3, MUL VL]\n" - "st1w z10.s, p0, [%[outptr3], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z11.s, p0, [%[outptr4], #1, MUL VL]\n" - "st1w z8.s, p0, [%[outptr5], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr6], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "add z13.s, z13.s, z2.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" - "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" - "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "add z14.s, z14.s, z3.s\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z15.s, z15.s, z4.s\n" + "ld1w z17.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z16.s, z16.s, z2.s\n" + "ld1w z18.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z19.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "add z17.s, z17.s, z3.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "add z18.s, z18.s, z4.s\n" + "ld1w z20.s, p1/z, [%[inptr], #7, MUL VL]\n" + "add z19.s, z19.s, z2.s\n" + "ld1w z13.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z14.s, p0/z, [x8, #-7, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "add z20.s, z20.s, z3.s\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z4.s\n" + "ld1w z15.s, p1/z, [x8, #-6, MUL VL]\n" + "add z14.s, z14.s, z2.s\n" "addvl %[outptr0], %[outptr0], #3\n" - "add z8.s, z8.s, z4.s\n" - "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "st1w z16.s, p0, [%[outptr1]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "add z15.s, z15.s, z3.s\n" + "ld1w z16.s, p2/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z17.s, p1, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "add z16.s, z16.s, z4.s\n" + "ld1w z17.s, p0/z, [x8, #-4, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z18.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z6.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "add z17.s, z17.s, z2.s\n" + "ld1w z18.s, p1/z, [x8, #-3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" - "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "st1w z19.s, p0, [%[outptr2]]\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "add z18.s, z18.s, z3.s\n" + "ld1w z19.s, p2/z, [x8, #-2, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "st1w z20.s, p1, [%[outptr2], #1, MUL VL]\n" + "ld1w z20.s, p0/z, [x8, #-1, MUL VL]\n" + "add z19.s, z19.s, z4.s\n" + "st1w z13.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "ld1w z7.s, p0/z, [%[outptr5], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #1, MUL VL]\n" + "add z20.s, z20.s, z2.s\n" + "ld1w z13.s, p1/z, [x8]\n" + "st1w z14.s, p0, [%[outptr3]]\n" + "ld1w z14.s, p2/z, [x8, #1, MUL VL]\n" + "add z13.s, z13.s, z3.s\n" + "st1w z15.s, p1, [%[outptr3], #1, MUL VL]\n" + "add z14.s, z14.s, z4.s\n" + "ld1w z15.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z16.s, p2, [%[outptr3], #2, MUL VL]\n" "addvl %[outptr3], %[outptr3], #3\n" - "ld1w z4.s, p0/z, [%[outptr6], #2, MUL VL]\n" - "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "add z15.s, z15.s, z2.s\n" + "ld1w z16.s, p1/z, [x8, #3, MUL VL]\n" + "st1w z17.s, p0, [%[outptr4]]\n" + "ld1w z17.s, p2/z, [x8, #4, MUL VL]\n" + "add z16.s, z16.s, z3.s\n" + "st1w z18.s, p1, [%[outptr4], #1, MUL VL]\n" + "add z17.s, z17.s, z4.s\n" + "st1w z19.s, p2, [%[outptr4], #2, MUL VL]\n" "addvl %[outptr4], %[outptr4], #3\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z8.s, p0/z, [x8, #4, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "st1w z11.s, p0, [%[outptr5], #2, MUL VL]\n" + "st1w z20.s, p0, [%[outptr5]]\n" + "st1w z13.s, p1, [%[outptr5], #1, MUL VL]\n" + "st1w z14.s, p2, [%[outptr5], #2, MUL VL]\n" "addvl %[outptr5], %[outptr5], #3\n" - "st1w z8.s, p0, [%[outptr6], #2, MUL VL]\n" + "st1w z15.s, p0, [%[outptr6]]\n" + "st1w z16.s, p1, [%[outptr6], #1, MUL VL]\n" + "st1w z17.s, p2, [%[outptr6], #2, MUL VL]\n" "addvl %[outptr6], %[outptr6], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; @@ -1199,141 +1280,114 @@ inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const in long p = 0; /* Optimized routine to copy an entire block */ __asm __volatile ( - "mov z2.s, %s[alpha]\n" "addvl x8, %[inptr], #16\n" - "mov z3.s, %s[beta]\n" "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" "incw %[p], all, mul #1\n" "prfm PLDL1KEEP, [%[inptr], #0x180]\n" - "ld1w z4.s, p0/z, [%[outptr0]]\n" - "prfm PLDL1KEEP, [%[inptr], #0x240]\n" - "ld1w z8.s, p0/z, [%[inptr]]\n" - "ld1w z5.s, p0/z, [%[outptr1]]\n" - "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3]]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr4]]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0]]\n" - "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z5.s, p0/z, [%[outptr5]]\n" - "ld1w z6.s, p0/z, [%[outptr6]]\n" - "st1w z9.s, p0, [%[outptr1]]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr7]]\n" - "st1w z10.s, p0, [%[outptr2]]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3]]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z11.s, p0/z, [x8, #5, MUL VL]\n" - "st1w z8.s, p0, [%[outptr4]]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z9.s, p0, [%[outptr5]]\n" - "st1w z10.s, p0, [%[outptr6]]\n" - "st1w z11.s, p0, [%[outptr7]]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z2.s, p0/z, [%[biasptr]]\n" + "whilelt p1.s, %[p], %[w]\n" + "ld1w z3.s, p0/z, [%[biasptr], #1, MUL VL]\n" "incw %[p], all, mul #1\n" + "ld1w z4.s, p0/z, [%[biasptr], #2, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" - "ld1w z4.s, p0/z, [%[outptr0], #1, MUL VL]\n" - "prfm PLDL1KEEP, [%[inptr], #0x280]\n" - "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" - "ld1w z5.s, p0/z, [%[outptr1], #1, MUL VL]\n" - "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2], #1, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" - "ld1w z4.s, p0/z, [%[outptr4], #1, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-3, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "ld1w z5.s, p0/z, [%[outptr5], #1, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr6], #1, MUL VL]\n" - "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [x8]\n" - "ld1w z7.s, p0/z, [%[outptr7], #1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" - "add z9.s, z9.s, z5.s\n" - "ld1w z10.s, p0/z, [x8, #3, MUL VL]\n" - "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z11.s, p0/z, [x8, #6, MUL VL]\n" - "st1w z8.s, p0, [%[outptr4], #1, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z9.s, p0, [%[outptr5], #1, MUL VL]\n" - "st1w z10.s, p0, [%[outptr6], #1, MUL VL]\n" - "st1w z11.s, p0, [%[outptr7], #1, MUL VL]\n" - "whilelt p0.s, %[p], %[w]\n" - "b.none 1f\n" - "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" - "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" - "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z13.s, p0/z, [%[inptr]]\n" + "whilelt p2.s, %[p], %[w]\n" + "ld1w z14.s, p1/z, [%[inptr], #1, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "add z13.s, z13.s, z2.s\n" + "ld1w z15.s, p2/z, [%[inptr], #2, MUL VL]\n" + "ld1w z16.s, p0/z, [%[inptr], #3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x200]\n" - "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" - "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" - "add z8.s, z8.s, z4.s\n" - "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" - "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" - "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" - "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "add z14.s, z14.s, z3.s\n" + "st1w z13.s, p0, [%[outptr0]]\n" + "add z15.s, z15.s, z4.s\n" + "ld1w z17.s, p1/z, [%[inptr], #4, MUL VL]\n" + "add z16.s, z16.s, z2.s\n" + "ld1w z18.s, p2/z, [%[inptr], #5, MUL VL]\n" + "ld1w z19.s, p0/z, [%[inptr], #6, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "add z17.s, z17.s, z3.s\n" + "st1w z14.s, p1, [%[outptr0], #1, MUL VL]\n" + "add z18.s, z18.s, z4.s\n" + "ld1w z20.s, p1/z, [%[inptr], #7, MUL VL]\n" + "add z19.s, z19.s, z2.s\n" + "ld1w z13.s, p2/z, [x8, #-8, MUL VL]\n" + "ld1w z14.s, p0/z, [x8, #-7, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "add z20.s, z20.s, z3.s\n" + "st1w z15.s, p2, [%[outptr0], #2, MUL VL]\n" + "add z13.s, z13.s, z4.s\n" + "ld1w z15.s, p1/z, [x8, #-6, MUL VL]\n" + "add z14.s, z14.s, z2.s\n" "addvl %[outptr0], %[outptr0], #3\n" - "add z10.s, z10.s, z6.s\n" - "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" - "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "st1w z16.s, p0, [%[outptr1]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "add z15.s, z15.s, z3.s\n" + "ld1w z16.s, p2/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z17.s, p1, [%[outptr1], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "add z16.s, z16.s, z4.s\n" + "ld1w z17.s, p0/z, [x8, #-4, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z18.s, p2, [%[outptr1], #2, MUL VL]\n" "addvl %[outptr1], %[outptr1], #3\n" - "ld1w z4.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "add z17.s, z17.s, z2.s\n" + "ld1w z18.s, p1/z, [x8, #-3, MUL VL]\n" "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" - "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "st1w z19.s, p0, [%[outptr2]]\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "add z18.s, z18.s, z3.s\n" + "ld1w z19.s, p2/z, [x8, #-2, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr7], #0x60]\n" + "st1w z20.s, p1, [%[outptr2], #1, MUL VL]\n" + "addvl %[inptr], %[inptr], #24\n" + "add z19.s, z19.s, z4.s\n" + "ld1w z20.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z13.s, p2, [%[outptr2], #2, MUL VL]\n" "addvl %[outptr2], %[outptr2], #3\n" - "ld1w z5.s, p0/z, [%[outptr5], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" - "add z8.s, z8.s, z4.s\n" - "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" - "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "add z20.s, z20.s, z2.s\n" + "ld1w z13.s, p1/z, [x8]\n" + "st1w z14.s, p0, [%[outptr3]]\n" + "ld1w z14.s, p2/z, [x8, #1, MUL VL]\n" + "add z13.s, z13.s, z3.s\n" + "st1w z15.s, p1, [%[outptr3], #1, MUL VL]\n" + "add z14.s, z14.s, z4.s\n" + "ld1w z15.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z16.s, p2, [%[outptr3], #2, MUL VL]\n" "addvl %[outptr3], %[outptr3], #3\n" - "ld1w z6.s, p0/z, [%[outptr6], #2, MUL VL]\n" - "prfm PLDL1KEEP, [%[outptr7], #0x60]\n" - "add z9.s, z9.s, z5.s\n" - "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" - "ld1w z10.s, p0/z, [x8, #4, MUL VL]\n" + "add z15.s, z15.s, z2.s\n" + "ld1w z16.s, p1/z, [x8, #3, MUL VL]\n" + "st1w z17.s, p0, [%[outptr4]]\n" + "ld1w z17.s, p2/z, [x8, #4, MUL VL]\n" + "add z16.s, z16.s, z3.s\n" + "st1w z18.s, p1, [%[outptr4], #1, MUL VL]\n" + "add z17.s, z17.s, z4.s\n" + "ld1w z18.s, p0/z, [x8, #5, MUL VL]\n" + "st1w z19.s, p2, [%[outptr4], #2, MUL VL]\n" "addvl %[outptr4], %[outptr4], #3\n" - "ld1w z7.s, p0/z, [%[outptr7], #2, MUL VL]\n" - "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "add z18.s, z18.s, z2.s\n" + "ld1w z19.s, p1/z, [x8, #6, MUL VL]\n" + "st1w z20.s, p0, [%[outptr5]]\n" + "ld1w z20.s, p2/z, [x8, #7, MUL VL]\n" + "add z19.s, z19.s, z3.s\n" + "st1w z13.s, p1, [%[outptr5], #1, MUL VL]\n" + "add z20.s, z20.s, z4.s\n" + "st1w z14.s, p2, [%[outptr5], #2, MUL VL]\n" "addvl %[outptr5], %[outptr5], #3\n" - "add z10.s, z10.s, z6.s\n" - "ld1w z11.s, p0/z, [x8, #7, MUL VL]\n" - "add z11.s, z11.s, z7.s\n" - "st1w z10.s, p0, [%[outptr6], #2, MUL VL]\n" + "st1w z15.s, p0, [%[outptr6]]\n" + "st1w z16.s, p1, [%[outptr6], #1, MUL VL]\n" + "st1w z17.s, p2, [%[outptr6], #2, MUL VL]\n" "addvl %[outptr6], %[outptr6], #3\n" - "st1w z11.s, p0, [%[outptr7], #2, MUL VL]\n" + "st1w z18.s, p0, [%[outptr7]]\n" + "st1w z19.s, p1, [%[outptr7], #1, MUL VL]\n" + "st1w z20.s, p2, [%[outptr7], #2, MUL VL]\n" "addvl %[outptr7], %[outptr7], #3\n" - "1:\n" - "addvl %[inptr], %[inptr], #24\n" : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), [inptr] "+r" (inptr), [p] "+r" (p) - : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) - : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + : [w] "r" (w), [biasptr] "r" (biasptr) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "memory", "cc" ); } break; |