mempool/octeontx2: add optimized dequeue operation for arm64
authorPavan Nikhilesh <pbhagavatula@marvell.com>
Sat, 22 Jun 2019 13:24:15 +0000 (18:54 +0530)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 25 Jun 2019 21:35:57 +0000 (23:35 +0200)
This patch adds an optimized arm64 instruction based routine to leverage
CPU pipeline characteristics of octeontx2. The theme is to fill the
pipeline with CASP operations as much HW can do so that HW can do alloc()
HW ops in full throttle.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Signed-off-by: Jerin Jacob <jerinj@marvell.com>
Signed-off-by: Vamsi Attunuru <vattunuru@marvell.com>
drivers/mempool/octeontx2/otx2_mempool_ops.c

index c59bd73..2517001 100644 (file)
@@ -37,6 +37,303 @@ npa_lf_aura_op_alloc_one(const int64_t wdata, int64_t * const addr,
        return -ENOENT;
 }
 
+#if defined(RTE_ARCH_ARM64)
+static __rte_noinline int
+npa_lf_aura_op_search_alloc(const int64_t wdata, int64_t * const addr,
+               void **obj_table, unsigned int n)
+{
+       uint8_t i;
+
+       for (i = 0; i < n; i++) {
+               if (obj_table[i] != NULL)
+                       continue;
+               if (npa_lf_aura_op_alloc_one(wdata, addr, obj_table, i))
+                       return -ENOENT;
+       }
+
+       return 0;
+}
+
+/*
+ * Some versions of the compiler don't have support for __int128_t for
+ * CASP inline-asm. i.e. if the optimization level is reduced to -O0 the
+ * CASP restrictions aren't followed and the compiler might end up violation the
+ * CASP rules. Fix it by explicitly providing ((optimize("-O3"))).
+ *
+ * Example:
+ * ccSPMGzq.s:1648: Error: reg pair must start from even reg at
+ * operand 1 - `casp x21,x22,x0,x1,[x19]'
+ */
+static  __attribute__((optimize("-O3"))) __rte_noinline int __hot
+npa_lf_aura_op_alloc_bulk(const int64_t wdata, int64_t * const addr,
+                         unsigned int n, void **obj_table)
+{
+       const __uint128_t wdata128 = ((__uint128_t)wdata << 64) | wdata;
+       uint64x2_t failed = vdupq_n_u64(~0);
+
+       switch (n) {
+       case 32:
+       {
+               __uint128_t t0, t1, t2, t3, t4, t5, t6, t7, t8, t9;
+               __uint128_t t10, t11;
+
+               asm volatile (
+               ".cpu  generic+lse\n"
+               "casp %[t0], %H[t0], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t1], %H[t1], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t2], %H[t2], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t3], %H[t3], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t4], %H[t4], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t5], %H[t5], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t6], %H[t6], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t7], %H[t7], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t8], %H[t8], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t9], %H[t9], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t10], %H[t10], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t11], %H[t11], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d16, %[t0]\n"
+               "fmov v16.D[1], %H[t0]\n"
+               "casp %[t0], %H[t0], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d17, %[t1]\n"
+               "fmov v17.D[1], %H[t1]\n"
+               "casp %[t1], %H[t1], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d18, %[t2]\n"
+               "fmov v18.D[1], %H[t2]\n"
+               "casp %[t2], %H[t2], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d19, %[t3]\n"
+               "fmov v19.D[1], %H[t3]\n"
+               "casp %[t3], %H[t3], %[wdata], %H[wdata], [%[loc]]\n"
+               "and %[failed].16B, %[failed].16B, v16.16B\n"
+               "and %[failed].16B, %[failed].16B, v17.16B\n"
+               "and %[failed].16B, %[failed].16B, v18.16B\n"
+               "and %[failed].16B, %[failed].16B, v19.16B\n"
+               "fmov d20, %[t4]\n"
+               "fmov v20.D[1], %H[t4]\n"
+               "fmov d21, %[t5]\n"
+               "fmov v21.D[1], %H[t5]\n"
+               "fmov d22, %[t6]\n"
+               "fmov v22.D[1], %H[t6]\n"
+               "fmov d23, %[t7]\n"
+               "fmov v23.D[1], %H[t7]\n"
+               "and %[failed].16B, %[failed].16B, v20.16B\n"
+               "and %[failed].16B, %[failed].16B, v21.16B\n"
+               "and %[failed].16B, %[failed].16B, v22.16B\n"
+               "and %[failed].16B, %[failed].16B, v23.16B\n"
+               "st1 { v16.2d, v17.2d, v18.2d, v19.2d}, [%[dst]], 64\n"
+               "st1 { v20.2d, v21.2d, v22.2d, v23.2d}, [%[dst]], 64\n"
+               "fmov d16, %[t8]\n"
+               "fmov v16.D[1], %H[t8]\n"
+               "fmov d17, %[t9]\n"
+               "fmov v17.D[1], %H[t9]\n"
+               "fmov d18, %[t10]\n"
+               "fmov v18.D[1], %H[t10]\n"
+               "fmov d19, %[t11]\n"
+               "fmov v19.D[1], %H[t11]\n"
+               "and %[failed].16B, %[failed].16B, v16.16B\n"
+               "and %[failed].16B, %[failed].16B, v17.16B\n"
+               "and %[failed].16B, %[failed].16B, v18.16B\n"
+               "and %[failed].16B, %[failed].16B, v19.16B\n"
+               "fmov d20, %[t0]\n"
+               "fmov v20.D[1], %H[t0]\n"
+               "fmov d21, %[t1]\n"
+               "fmov v21.D[1], %H[t1]\n"
+               "fmov d22, %[t2]\n"
+               "fmov v22.D[1], %H[t2]\n"
+               "fmov d23, %[t3]\n"
+               "fmov v23.D[1], %H[t3]\n"
+               "and %[failed].16B, %[failed].16B, v20.16B\n"
+               "and %[failed].16B, %[failed].16B, v21.16B\n"
+               "and %[failed].16B, %[failed].16B, v22.16B\n"
+               "and %[failed].16B, %[failed].16B, v23.16B\n"
+               "st1 { v16.2d, v17.2d, v18.2d, v19.2d}, [%[dst]], 64\n"
+               "st1 { v20.2d, v21.2d, v22.2d, v23.2d}, [%[dst]], 64\n"
+               : "+Q" (*addr), [failed] "=&w" (failed),
+               [t0] "=&r" (t0), [t1] "=&r" (t1), [t2] "=&r" (t2),
+               [t3] "=&r" (t3), [t4] "=&r" (t4), [t5] "=&r" (t5),
+               [t6] "=&r" (t6), [t7] "=&r" (t7), [t8] "=&r" (t8),
+               [t9] "=&r" (t9), [t10] "=&r" (t10), [t11] "=&r" (t11)
+               : [wdata] "r" (wdata128), [dst] "r" (obj_table),
+               [loc] "r" (addr)
+               : "memory", "v16", "v17", "v18",
+               "v19", "v20", "v21", "v22", "v23"
+               );
+               break;
+       }
+       case 16:
+       {
+               __uint128_t t0, t1, t2, t3, t4, t5, t6, t7;
+
+               asm volatile (
+               ".cpu  generic+lse\n"
+               "casp %[t0], %H[t0], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t1], %H[t1], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t2], %H[t2], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t3], %H[t3], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t4], %H[t4], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t5], %H[t5], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t6], %H[t6], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t7], %H[t7], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d16, %[t0]\n"
+               "fmov v16.D[1], %H[t0]\n"
+               "fmov d17, %[t1]\n"
+               "fmov v17.D[1], %H[t1]\n"
+               "fmov d18, %[t2]\n"
+               "fmov v18.D[1], %H[t2]\n"
+               "fmov d19, %[t3]\n"
+               "fmov v19.D[1], %H[t3]\n"
+               "and %[failed].16B, %[failed].16B, v16.16B\n"
+               "and %[failed].16B, %[failed].16B, v17.16B\n"
+               "and %[failed].16B, %[failed].16B, v18.16B\n"
+               "and %[failed].16B, %[failed].16B, v19.16B\n"
+               "fmov d20, %[t4]\n"
+               "fmov v20.D[1], %H[t4]\n"
+               "fmov d21, %[t5]\n"
+               "fmov v21.D[1], %H[t5]\n"
+               "fmov d22, %[t6]\n"
+               "fmov v22.D[1], %H[t6]\n"
+               "fmov d23, %[t7]\n"
+               "fmov v23.D[1], %H[t7]\n"
+               "and %[failed].16B, %[failed].16B, v20.16B\n"
+               "and %[failed].16B, %[failed].16B, v21.16B\n"
+               "and %[failed].16B, %[failed].16B, v22.16B\n"
+               "and %[failed].16B, %[failed].16B, v23.16B\n"
+               "st1 { v16.2d, v17.2d, v18.2d, v19.2d}, [%[dst]], 64\n"
+               "st1 { v20.2d, v21.2d, v22.2d, v23.2d}, [%[dst]], 64\n"
+               : "+Q" (*addr), [failed] "=&w" (failed),
+               [t0] "=&r" (t0), [t1] "=&r" (t1), [t2] "=&r" (t2),
+               [t3] "=&r" (t3), [t4] "=&r" (t4), [t5] "=&r" (t5),
+               [t6] "=&r" (t6), [t7] "=&r" (t7)
+               : [wdata] "r" (wdata128), [dst] "r" (obj_table),
+               [loc] "r" (addr)
+               : "memory", "v16", "v17", "v18", "v19",
+                 "v20", "v21", "v22", "v23"
+               );
+               break;
+       }
+       case 8:
+       {
+               __uint128_t t0, t1, t2, t3;
+
+               asm volatile (
+               ".cpu  generic+lse\n"
+               "casp %[t0], %H[t0], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t1], %H[t1], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t2], %H[t2], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t3], %H[t3], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d16, %[t0]\n"
+               "fmov v16.D[1], %H[t0]\n"
+               "fmov d17, %[t1]\n"
+               "fmov v17.D[1], %H[t1]\n"
+               "fmov d18, %[t2]\n"
+               "fmov v18.D[1], %H[t2]\n"
+               "fmov d19, %[t3]\n"
+               "fmov v19.D[1], %H[t3]\n"
+               "and %[failed].16B, %[failed].16B, v16.16B\n"
+               "and %[failed].16B, %[failed].16B, v17.16B\n"
+               "and %[failed].16B, %[failed].16B, v18.16B\n"
+               "and %[failed].16B, %[failed].16B, v19.16B\n"
+               "st1 { v16.2d, v17.2d, v18.2d, v19.2d}, [%[dst]], 64\n"
+               : "+Q" (*addr), [failed] "=&w" (failed),
+               [t0] "=&r" (t0), [t1] "=&r" (t1), [t2] "=&r" (t2),
+               [t3] "=&r" (t3)
+               : [wdata] "r" (wdata128), [dst] "r" (obj_table),
+               [loc] "r" (addr)
+               : "memory", "v16", "v17", "v18", "v19"
+               );
+               break;
+       }
+       case 4:
+       {
+               __uint128_t t0, t1;
+
+               asm volatile (
+               ".cpu  generic+lse\n"
+               "casp %[t0], %H[t0], %[wdata], %H[wdata], [%[loc]]\n"
+               "casp %[t1], %H[t1], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d16, %[t0]\n"
+               "fmov v16.D[1], %H[t0]\n"
+               "fmov d17, %[t1]\n"
+               "fmov v17.D[1], %H[t1]\n"
+               "and %[failed].16B, %[failed].16B, v16.16B\n"
+               "and %[failed].16B, %[failed].16B, v17.16B\n"
+               "st1 { v16.2d, v17.2d}, [%[dst]], 32\n"
+               : "+Q" (*addr), [failed] "=&w" (failed),
+               [t0] "=&r" (t0), [t1] "=&r" (t1)
+               : [wdata] "r" (wdata128), [dst] "r" (obj_table),
+               [loc] "r" (addr)
+               : "memory", "v16", "v17"
+               );
+               break;
+       }
+       case 2:
+       {
+               __uint128_t t0;
+
+               asm volatile (
+               ".cpu  generic+lse\n"
+               "casp %[t0], %H[t0], %[wdata], %H[wdata], [%[loc]]\n"
+               "fmov d16, %[t0]\n"
+               "fmov v16.D[1], %H[t0]\n"
+               "and %[failed].16B, %[failed].16B, v16.16B\n"
+               "st1 { v16.2d}, [%[dst]], 16\n"
+               : "+Q" (*addr), [failed] "=&w" (failed),
+               [t0] "=&r" (t0)
+               : [wdata] "r" (wdata128), [dst] "r" (obj_table),
+               [loc] "r" (addr)
+               : "memory", "v16"
+               );
+               break;
+       }
+       case 1:
+               return npa_lf_aura_op_alloc_one(wdata, addr, obj_table, 0);
+       }
+
+       if (unlikely(!(vgetq_lane_u64(failed, 0) & vgetq_lane_u64(failed, 1))))
+               return npa_lf_aura_op_search_alloc(wdata, addr, (void **)
+                       ((char *)obj_table - (sizeof(uint64_t) * n)), n);
+
+       return 0;
+}
+
+static __rte_noinline void
+otx2_npa_clear_alloc(struct rte_mempool *mp, void **obj_table, unsigned int n)
+{
+       unsigned int i;
+
+       for (i = 0; i < n; i++) {
+               if (obj_table[i] != NULL) {
+                       otx2_npa_enq(mp, &obj_table[i], 1);
+                       obj_table[i] = NULL;
+               }
+       }
+}
+
+static inline int __hot
+otx2_npa_deq_arm64(struct rte_mempool *mp, void **obj_table, unsigned int n)
+{
+       const int64_t wdata = npa_lf_aura_handle_to_aura(mp->pool_id);
+       void **obj_table_bak = obj_table;
+       const unsigned int nfree = n;
+       unsigned int parts;
+
+       int64_t * const addr = (int64_t * const)
+                       (npa_lf_aura_handle_to_base(mp->pool_id) +
+                               NPA_LF_AURA_OP_ALLOCX(0));
+       while (n) {
+               parts = n > 31 ? 32 : rte_align32prevpow2(n);
+               n -= parts;
+               if (unlikely(npa_lf_aura_op_alloc_bulk(wdata, addr,
+                               parts, obj_table))) {
+                       otx2_npa_clear_alloc(mp, obj_table_bak, nfree - n);
+                       return -ENOENT;
+               }
+               obj_table += parts;
+       }
+
+       return 0;
+}
+#endif
+
 static inline int __hot
 otx2_npa_deq(struct rte_mempool *mp, void **obj_table, unsigned int n)
 {
@@ -463,7 +760,11 @@ static struct rte_mempool_ops otx2_npa_ops = {
        .get_count = otx2_npa_get_count,
        .calc_mem_size = otx2_npa_calc_mem_size,
        .populate = otx2_npa_populate,
+#if defined(RTE_ARCH_ARM64)
+       .dequeue = otx2_npa_deq_arm64,
+#else
        .dequeue = otx2_npa_deq,
+#endif
 };
 
 MEMPOOL_REGISTER_OPS(otx2_npa_ops);