diff mbox series

[RFC,v4,4/5] target/riscv: rvv: Provide group continuous ld/st flow for unit-stride ld/st instructions

Message ID 20240613175122.1299212-5-max.chou@sifive.com
State New
Headers show
Series Improve the performance of RISC-V vector unit-stride/whole register ld/st instructions | expand

Commit Message

Max Chou June 13, 2024, 5:51 p.m. UTC 
The vector unmasked unit-stride and whole register load/store
instructions will load/store continuous memory. If the endian of both
the host and guest architecture are the same, then we can group the
element load/store to load/store more data at a time.

Signed-off-by: Max Chou <max.chou@sifive.com>
---
 target/riscv/vector_helper.c | 160 +++++++++++++++++++++++++----------
 1 file changed, 117 insertions(+), 43 deletions(-)

Comments

Richard Henderson June 20, 2024, 4:38 a.m. UTC | #1 
On 6/13/24 10:51, Max Chou wrote:
> The vector unmasked unit-stride and whole register load/store
> instructions will load/store continuous memory. If the endian of both
> the host and guest architecture are the same, then we can group the
> element load/store to load/store more data at a time.
> 
> Signed-off-by: Max Chou <max.chou@sifive.com>
> ---
>   target/riscv/vector_helper.c | 160 +++++++++++++++++++++++++----------
>   1 file changed, 117 insertions(+), 43 deletions(-)
> 
> diff --git a/target/riscv/vector_helper.c b/target/riscv/vector_helper.c
> index 793337a6f96..cba46ef16a5 100644
> --- a/target/riscv/vector_helper.c
> +++ b/target/riscv/vector_helper.c
> @@ -457,6 +457,69 @@ GEN_VEXT_ST_ELEM(ste_h, uint16_t, H2, stw)
>   GEN_VEXT_ST_ELEM(ste_w, uint32_t, H4, stl)
>   GEN_VEXT_ST_ELEM(ste_d, uint64_t, H8, stq)
>   
> +static inline uint32_t
> +vext_group_ldst_host(CPURISCVState *env, void *vd, uint32_t byte_end,
> +                     uint32_t byte_offset, void *host, uint32_t esz,
> +                     bool is_load)
> +{
> +    uint32_t group_size;
> +    static vext_ldst_elem_fn_host * const fns[2][4] = {
> +        /* Store */
> +        { ste_b_host, ste_h_host, ste_w_host, ste_d_host },
> +        /* Load */
> +        { lde_b_host, lde_h_host, lde_w_host, lde_d_host }
> +    };
> +    vext_ldst_elem_fn_host *fn;
> +
> +    if (byte_offset + 8 < byte_end) {
> +        group_size = MO_64;
> +    } else if (byte_offset + 4 < byte_end) {
> +        group_size = MO_32;
> +    } else if (byte_offset + 2 < byte_end) {
> +        group_size = MO_16;
> +    } else {
> +        group_size = MO_8;
> +    }
> +
> +    fn = fns[is_load][group_size];
> +    fn(vd, byte_offset, host + byte_offset);

This is a really bad idea.  The table and indirect call means that none of these will be 
properly inlined.  Anyway...

> +
> +    return 1 << group_size;
> +}
> +
> +static inline void
> +vext_continus_ldst_tlb(CPURISCVState *env, vext_ldst_elem_fn_tlb *ldst_tlb,
> +                       void *vd, uint32_t evl, target_ulong addr,
> +                       uint32_t reg_start, uintptr_t ra, uint32_t esz,
> +                       bool is_load)
> +{
> +    for (; reg_start < evl; reg_start++, addr += esz) {
> +        ldst_tlb(env, adjust_addr(env, addr), reg_start * esz, vd, ra);
> +    }
> +}
> +
> +static inline void
> +vext_continus_ldst_host(CPURISCVState *env, vext_ldst_elem_fn_host *ldst_host,
> +                        void *vd, uint32_t evl, uint32_t reg_start, void *host,
> +                        uint32_t esz, bool is_load)
> +{
> +#if TARGET_BIG_ENDIAN != HOST_BIG_ENDIAN
> +    for (; reg_start < evl; reg_start++) {
> +        uint32_t byte_off = reg_start * esz;
> +        ldst_host(vd, byte_off, host + byte_off);
> +    }
> +#else
> +    uint32_t group_byte;
> +    uint32_t byte_start = reg_start * esz;
> +    uint32_t byte_end = evl * esz;
> +    while (byte_start < byte_end) {
> +        group_byte = vext_group_ldst_host(env, vd, byte_end, byte_start, host,
> +                                          esz, is_load);
> +        byte_start += group_byte;
> +    }

... this is much better handled with memcpy, given that you know endianness matches.


r~
Max Chou June 24, 2024, 6:50 a.m. UTC | #2 
On 2024/6/20 12:38 PM, Richard Henderson wrote:

> On 6/13/24 10:51, Max Chou wrote:
>> The vector unmasked unit-stride and whole register load/store
>> instructions will load/store continuous memory. If the endian of both
>> the host and guest architecture are the same, then we can group the
>> element load/store to load/store more data at a time.
>>
>> Signed-off-by: Max Chou <max.chou@sifive.com>
>> ---
>>   target/riscv/vector_helper.c | 160 +++++++++++++++++++++++++----------
>>   1 file changed, 117 insertions(+), 43 deletions(-)
>>
>> diff --git a/target/riscv/vector_helper.c b/target/riscv/vector_helper.c
>> index 793337a6f96..cba46ef16a5 100644
>> --- a/target/riscv/vector_helper.c
>> +++ b/target/riscv/vector_helper.c
>> @@ -457,6 +457,69 @@ GEN_VEXT_ST_ELEM(ste_h, uint16_t, H2, stw)
>>   GEN_VEXT_ST_ELEM(ste_w, uint32_t, H4, stl)
>>   GEN_VEXT_ST_ELEM(ste_d, uint64_t, H8, stq)
>>   +static inline uint32_t
>> +vext_group_ldst_host(CPURISCVState *env, void *vd, uint32_t byte_end,
>> +                     uint32_t byte_offset, void *host, uint32_t esz,
>> +                     bool is_load)
>> +{
>> +    uint32_t group_size;
>> +    static vext_ldst_elem_fn_host * const fns[2][4] = {
>> +        /* Store */
>> +        { ste_b_host, ste_h_host, ste_w_host, ste_d_host },
>> +        /* Load */
>> +        { lde_b_host, lde_h_host, lde_w_host, lde_d_host }
>> +    };
>> +    vext_ldst_elem_fn_host *fn;
>> +
>> +    if (byte_offset + 8 < byte_end) {
>> +        group_size = MO_64;
>> +    } else if (byte_offset + 4 < byte_end) {
>> +        group_size = MO_32;
>> +    } else if (byte_offset + 2 < byte_end) {
>> +        group_size = MO_16;
>> +    } else {
>> +        group_size = MO_8;
>> +    }
>> +
>> +    fn = fns[is_load][group_size];
>> +    fn(vd, byte_offset, host + byte_offset);
>
> This is a really bad idea.  The table and indirect call means that 
> none of these will be properly inlined.  Anyway...
>
>> +
>> +    return 1 << group_size;
>> +}
>> +
>> +static inline void
>> +vext_continus_ldst_tlb(CPURISCVState *env, vext_ldst_elem_fn_tlb 
>> *ldst_tlb,
>> +                       void *vd, uint32_t evl, target_ulong addr,
>> +                       uint32_t reg_start, uintptr_t ra, uint32_t esz,
>> +                       bool is_load)
>> +{
>> +    for (; reg_start < evl; reg_start++, addr += esz) {
>> +        ldst_tlb(env, adjust_addr(env, addr), reg_start * esz, vd, ra);
>> +    }
>> +}
>> +
>> +static inline void
>> +vext_continus_ldst_host(CPURISCVState *env, vext_ldst_elem_fn_host 
>> *ldst_host,
>> +                        void *vd, uint32_t evl, uint32_t reg_start, 
>> void *host,
>> +                        uint32_t esz, bool is_load)
>> +{
>> +#if TARGET_BIG_ENDIAN != HOST_BIG_ENDIAN
>> +    for (; reg_start < evl; reg_start++) {
>> +        uint32_t byte_off = reg_start * esz;
>> +        ldst_host(vd, byte_off, host + byte_off);
>> +    }
>> +#else
>> +    uint32_t group_byte;
>> +    uint32_t byte_start = reg_start * esz;
>> +    uint32_t byte_end = evl * esz;
>> +    while (byte_start < byte_end) {
>> +        group_byte = vext_group_ldst_host(env, vd, byte_end, 
>> byte_start, host,
>> +                                          esz, is_load);
>> +        byte_start += group_byte;
>> +    }
>
> ... this is much better handled with memcpy, given that you know 
> endianness matches.
Thanks for the suggestion.
I'll try to replace the original implementation of the table and 
indirect calls by handled with memcpy at the next version.

Max.
>
>
> r~
diff mbox series

Patch

diff --git a/target/riscv/vector_helper.c b/target/riscv/vector_helper.c
index 793337a6f96..cba46ef16a5 100644
--- a/target/riscv/vector_helper.c
+++ b/target/riscv/vector_helper.c
@@ -457,6 +457,69 @@  GEN_VEXT_ST_ELEM(ste_h, uint16_t, H2, stw)
 GEN_VEXT_ST_ELEM(ste_w, uint32_t, H4, stl)
 GEN_VEXT_ST_ELEM(ste_d, uint64_t, H8, stq)
 
+static inline uint32_t
+vext_group_ldst_host(CPURISCVState *env, void *vd, uint32_t byte_end,
+                     uint32_t byte_offset, void *host, uint32_t esz,
+                     bool is_load)
+{
+    uint32_t group_size;
+    static vext_ldst_elem_fn_host * const fns[2][4] = {
+        /* Store */
+        { ste_b_host, ste_h_host, ste_w_host, ste_d_host },
+        /* Load */
+        { lde_b_host, lde_h_host, lde_w_host, lde_d_host }
+    };
+    vext_ldst_elem_fn_host *fn;
+
+    if (byte_offset + 8 < byte_end) {
+        group_size = MO_64;
+    } else if (byte_offset + 4 < byte_end) {
+        group_size = MO_32;
+    } else if (byte_offset + 2 < byte_end) {
+        group_size = MO_16;
+    } else {
+        group_size = MO_8;
+    }
+
+    fn = fns[is_load][group_size];
+    fn(vd, byte_offset, host + byte_offset);
+
+    return 1 << group_size;
+}
+
+static inline void
+vext_continus_ldst_tlb(CPURISCVState *env, vext_ldst_elem_fn_tlb *ldst_tlb,
+                       void *vd, uint32_t evl, target_ulong addr,
+                       uint32_t reg_start, uintptr_t ra, uint32_t esz,
+                       bool is_load)
+{
+    for (; reg_start < evl; reg_start++, addr += esz) {
+        ldst_tlb(env, adjust_addr(env, addr), reg_start * esz, vd, ra);
+    }
+}
+
+static inline void
+vext_continus_ldst_host(CPURISCVState *env, vext_ldst_elem_fn_host *ldst_host,
+                        void *vd, uint32_t evl, uint32_t reg_start, void *host,
+                        uint32_t esz, bool is_load)
+{
+#if TARGET_BIG_ENDIAN != HOST_BIG_ENDIAN
+    for (; reg_start < evl; reg_start++) {
+        uint32_t byte_off = reg_start * esz;
+        ldst_host(vd, byte_off, host + byte_off);
+    }
+#else
+    uint32_t group_byte;
+    uint32_t byte_start = reg_start * esz;
+    uint32_t byte_end = evl * esz;
+    while (byte_start < byte_end) {
+        group_byte = vext_group_ldst_host(env, vd, byte_end, byte_start, host,
+                                          esz, is_load);
+        byte_start += group_byte;
+    }
+#endif
+}
+
 static void vext_set_tail_elems_1s(target_ulong vl, void *vd,
                                    uint32_t desc, uint32_t nf,
                                    uint32_t esz, uint32_t max_elems)
@@ -555,6 +618,7 @@  vext_ldst_us(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
              uint32_t evl, uintptr_t ra, bool is_load)
 {
     RVVContLdSt info;
+    target_ulong addr;
     void *host;
     int flags;
     intptr_t reg_start, reg_last;
@@ -587,13 +651,19 @@  vext_ldst_us(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
         }
         reg_last += 1;
 
-        for (i = reg_start; i < reg_last; ++i) {
-            k = 0;
-            while (k < nf) {
-                target_ulong addr = base + ((i * nf + k) << log2_esz);
-                ldst_tlb(env, adjust_addr(env, addr),
-                         (i + k * max_elems) << log2_esz, vd, ra);
-                k++;
+        if (nf == 1) {
+            addr = base + reg_start * esz;
+            vext_continus_ldst_tlb(env, ldst_tlb, vd, reg_last, addr,
+                                   reg_start, ra, esz, is_load);
+        } else {
+            for (i = reg_start; i < reg_last; ++i) {
+                k = 0;
+                while (k < nf) {
+                    addr = base + ((i * nf + k) * esz);
+                    ldst_tlb(env, adjust_addr(env, addr),
+                             (i + k * max_elems) << log2_esz, vd, ra);
+                    k++;
+                }
             }
         }
 
@@ -607,12 +677,17 @@  vext_ldst_us(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
     reg_last = info.reg_idx_last[0] + 1;
     host = info.page[0].host;
 
-    for (i = reg_start; i < reg_last; ++i) {
-        k = 0;
-        while (k < nf) {
-            ldst_host(vd, (i + k * max_elems) << log2_esz,
-                      host + ((i * nf + k) << log2_esz));
-            k++;
+    if (nf == 1) {
+        vext_continus_ldst_host(env, ldst_host, vd, reg_last, reg_start, host,
+                                esz, is_load);
+    } else {
+        for (i = reg_start; i < reg_last; ++i) {
+            k = 0;
+            while (k < nf) {
+                ldst_host(vd, (i + k * max_elems) << log2_esz,
+                          host + ((i * nf + k) * esz));
+                k++;
+            }
         }
     }
 
@@ -624,7 +699,7 @@  vext_ldst_us(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
         reg_start = info.reg_idx_split;
         k = 0;
         while (k < nf) {
-            target_ulong addr = base + ((reg_start * nf + k) << log2_esz);
+            addr = base + ((reg_start * nf + k) << log2_esz);
             ldst_tlb(env, adjust_addr(env, addr),
                      (reg_start + k * max_elems) << log2_esz, vd, ra);
             k++;
@@ -636,12 +711,17 @@  vext_ldst_us(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
         reg_last = info.reg_idx_last[1] + 1;
         host = info.page[1].host;
 
-        for (i = reg_start; i < reg_last; ++i) {
-            k = 0;
-            while (k < nf) {
-                ldst_host(vd, (i + k * max_elems) << log2_esz,
-                          host + ((i * nf + k) << log2_esz));
-                k++;
+        if (nf == 1) {
+            vext_continus_ldst_host(env, ldst_host, vd, reg_last, reg_start,
+                                    host, esz, is_load);
+        } else {
+            for (i = reg_start; i < reg_last; ++i) {
+                k = 0;
+                while (k < nf) {
+                    ldst_host(vd, (i + k * max_elems) << log2_esz,
+                              host + ((i * nf + k) << log2_esz));
+                    k++;
+                }
             }
         }
     }
@@ -974,20 +1054,17 @@  vext_ldst_whole(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
              * load/store rest of elements of current segment pointed by vstart
              */
             addr = base + (reg_start << log2_esz);
-            for (; reg_start < evl; reg_start++, addr += esz) {
-                ldst_tlb(env, adjust_addr(env, addr), reg_start << log2_esz,
-                         vd, ra);
-            }
+            vext_continus_ldst_tlb(env, ldst_tlb, vd, evl, addr, reg_start, ra,
+                                   esz, is_load);
             idx_nf++;
         }
 
         /* load/store elements for rest of segments */
         evl = nf * max_elems;
         addr = base + (reg_start << log2_esz);
-        for (; reg_start < evl; reg_start++, addr += esz) {
-            ldst_tlb(env, adjust_addr(env, addr), reg_start << log2_esz, vd,
-                     ra);
-        }
+        reg_start = idx_nf * max_elems;
+        vext_continus_ldst_tlb(env, ldst_tlb, vd, evl, addr, reg_start, ra,
+                               esz, is_load);
 
         env->vstart = 0;
         return;
@@ -1003,17 +1080,16 @@  vext_ldst_whole(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
 
     if (off) {
         /* load/store rest of elements of current segment pointed by vstart */
-        for (; reg_start < evl; reg_start++) {
-            ldst_host(vd, reg_start << log2_esz,
-                      host + (reg_start << log2_esz));
-        }
+        vext_continus_ldst_host(env, ldst_host, vd, evl, reg_start, host, esz,
+                                is_load);
         idx_nf++;
     }
 
     /* load/store elements for rest of segments */
-    for (; reg_start < reg_last; reg_start++) {
-        ldst_host(vd, reg_start << log2_esz, host + (reg_start << log2_esz));
-    }
+    evl = reg_last;
+    reg_start = idx_nf * max_elems;
+    vext_continus_ldst_host(env, ldst_host, vd, evl, reg_start, host, esz,
+                            is_load);
 
     /*
      * Use the slow path to manage the cross-page misalignment.
@@ -1037,18 +1113,16 @@  vext_ldst_whole(void *vd, target_ulong base, CPURISCVState *env, uint32_t desc,
             /*
              * load/store rest of elements of current segment pointed by vstart
              */
-            for (; reg_start < evl; reg_start++) {
-                ldst_host(vd, reg_start << log2_esz,
-                          host + (reg_start << log2_esz));
-            }
+            vext_continus_ldst_host(env, ldst_host, vd, evl, reg_start, host,
+                                    esz, is_load);
             idx_nf++;
         }
 
         /* load/store elements for rest of segments */
-        for (; reg_start < reg_last; reg_start++) {
-            ldst_host(vd, reg_start << log2_esz,
-                      host + (reg_start << log2_esz));
-        }
+        evl = reg_last;
+        reg_start = idx_nf * max_elems;
+        vext_continus_ldst_host(env, ldst_host, vd, evl, reg_start, host, esz,
+                                is_load);
     }
 
     env->vstart = 0;