Skip to contentSkip to conversation

target/arm: Implement SVE2 fp conversion

RFC PATCH Applied Stephen Long 2 participants 6 messages Apr 28, 2020 – Jun 16, 2020

lore
Share & tools
lore
#0 lore raw

cover target/arm: Implement SVE2 fp conversion

The only insn left in this category is BFCVTNT. I couldn't find any helper functions for bfloats so I'm not sure how to approach this insn.

Also, for insn FCVTXNT, I'm not sure if it is ok to set the status rounding mode to odd like that.

Stephen Long (3):
target/arm: Implement SVE2 FCVTNT
target/arm: Implement SVE2 FCVTLT
target/arm: Implement SVE2 FCVTXNT

 target/arm/helper-sve.h    | 12 +++++++++
 target/arm/sve.decode      |  7 ++++++
 target/arm/sve_helper.c    | 50 ++++++++++++++++++++++++++++++++++++++
 target/arm/translate-sve.c | 40 ++++++++++++++++++++++++++++++
 4 files changed, 109 insertions(+)
Signature · 1 line
2.17.1
#1 lore raw

RFC PATCH 1/3 target/arm: Implement SVE2 FCVTNT

---
 target/arm/helper-sve.h    |  5 +++++
 target/arm/sve.decode      |  4 ++++
 target/arm/sve_helper.c    | 20 ++++++++++++++++++++
 target/arm/translate-sve.c | 16 ++++++++++++++++
 4 files changed, 45 insertions(+)
target/arm/helper-sve.h +5 -0
index 6b64c844bf..154593f110 100644
@@ -2713,3 +2713,8 @@ DEF_HELPER_FLAGS_5(sve2_sqdmlsl_idx_s, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(sve2_sqdmlsl_idx_d, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_5(sve2_fcvtnt_sh, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(sve2_fcvtnt_ds, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
target/arm/sve.decode +4 -0
index 22ed626784..355b8eec75 100644
@@ -1478,3 +1478,7 @@ STNT1_zprz      1110010 .. 00 ..... 001 ... ..... ..... \
 # SVE2 32-bit scatter non-temporal store (vector plus scalar)
 STNT1_zprz      1110010 .. 10 ..... 001 ... ..... ..... \
                 @rprr_scatter_store xs=0 esz=2 scale=0
+
+### SVE2 floating-point convert precision odd elements
+FCVTNT_sh       01100100 10 0010 00 101 ... ..... .....         @rd_pg_rn_e0
+FCVTNT_ds       01100100 11 0010 10 101 ... ..... .....         @rd_pg_rn_e0
target/arm/sve_helper.c +20 -0
index b8c8d10c99..a35983b7cd 100644
@@ -7570,3 +7570,23 @@ void HELPER(fmmla_d)(void *vd, void *va, void *vn, void *vm,
         d[3] = float64_add(a[3], float64_add(p0, p1, status), status);
     }
 }
+
+#define DO_FCVTNT(NAME, TYPEW, TYPEN, HW, HN, OP)                             \
+void HELPER(NAME)(void *vd, void *vn, void *vg, void *status, uint32_t desc)  \
+{                                                                             \
+    intptr_t i = simd_oprsz(desc);                                            \
+    uint64_t *g = vg;                                                         \
+    do {                                                                      \
+        uint64_t pg = g[(i - 1) >> 6];                                        \
+        do {                                                                  \
+            i -= sizeof(TYPEW);                                               \
+            if (likely((pg >> (i & 63)) & 1)) {                               \
+                TYPEW nn = *(TYPEW *)(vn + HW(i));                            \
+                *(TYPEN *)(vd + HN(i + sizeof(TYPEN))) = OP(nn, status);      \
+            }                                                                 \
+        } while (i & 63);                                                     \
+    } while (i != 0);                                                         \
+}
+
+DO_FCVTNT(sve2_fcvtnt_sh, uint32_t, uint16_t, H1_4, H1_2, sve_f32_to_f16)
+DO_FCVTNT(sve2_fcvtnt_ds, uint64_t, uint32_t, H1_4, H1_2, float64_to_float32)
target/arm/translate-sve.c +16 -0
index 5303d2e236..d410d0024e 100644
@@ -8105,3 +8105,19 @@ static bool trans_SQRDCMLAH_zzzz(DisasContext *s, arg_CMLA_zzzz *a)
     };
     return do_sve2_zzzz_fn(s, a->rd, a->rn, a->rm, a->ra, fns[a->esz], a->rot);
 }
+
+static bool trans_FCVTNT_sh(DisasContext *s, arg_rpr_esz *a)
+{
+    if (!dc_isar_feature(aa64_sve2, s)) {
+        return false;
+    }
+    return do_zpz_ptr(s, a->rd, a->rn, a->pg, false, gen_helper_sve2_fcvtnt_sh);
+}
+
+static bool trans_FCVTNT_ds(DisasContext *s, arg_rpr_esz *a)
+{
+    if (!dc_isar_feature(aa64_sve2, s)) {
+        return false;
+    }
+    return do_zpz_ptr(s, a->rd, a->rn, a->pg, false, gen_helper_sve2_fcvtnt_ds);
+}
Signature · 1 line
2.17.1
#2 lore raw

RFC PATCH 2/3 target/arm: Implement SVE2 FCVTLT

In reply to Stephen Long
---
 target/arm/helper-sve.h    |  5 +++++
 target/arm/sve.decode      |  2 ++
 target/arm/sve_helper.c    | 23 +++++++++++++++++++++++
 target/arm/translate-sve.c | 16 ++++++++++++++++
 4 files changed, 46 insertions(+)
target/arm/helper-sve.h +5 -0
index 154593f110..27a8cc2cbe 100644
@@ -2718,3 +2718,8 @@ DEF_HELPER_FLAGS_5(sve2_fcvtnt_sh, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(sve2_fcvtnt_ds, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
+
+DEF_HELPER_FLAGS_5(sve2_fcvtlt_hs, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(sve2_fcvtlt_sd, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
target/arm/sve.decode +2 -0
index 355b8eec75..95fc7ee32c 100644
@@ -1481,4 +1481,6 @@ STNT1_zprz      1110010 .. 10 ..... 001 ... ..... ..... \
 
 ### SVE2 floating-point convert precision odd elements
 FCVTNT_sh       01100100 10 0010 00 101 ... ..... .....         @rd_pg_rn_e0
+FCVTLT_hs       01100100 10 0010 01 101 ... ..... .....         @rd_pg_rn_e0
 FCVTNT_ds       01100100 11 0010 10 101 ... ..... .....         @rd_pg_rn_e0
+FCVTLT_sd       01100100 11 0010 11 101 ... ..... .....         @rd_pg_rn_e0
target/arm/sve_helper.c +23 -0
index a35983b7cd..8461d84445 100644
@@ -7590,3 +7590,26 @@ void HELPER(NAME)(void *vd, void *vn, void *vg, void *status, uint32_t desc)  \
 
 DO_FCVTNT(sve2_fcvtnt_sh, uint32_t, uint16_t, H1_4, H1_2, sve_f32_to_f16)
 DO_FCVTNT(sve2_fcvtnt_ds, uint64_t, uint32_t, H1_4, H1_2, float64_to_float32)
+
+#define DO_FCVTLT(NAME, TYPEW, TYPEN, HW, HN, OP)                             \
+void HELPER(NAME)(void *vd, void *vn, void *vg, void *status, uint32_t desc)  \
+{                                                                             \
+    intptr_t i = simd_oprsz(desc);                                            \
+    uint64_t *g = vg;                                                         \
+    do {                                                                      \
+        uint64_t pg = g[(i - 1) >> 6];                                        \
+        do {                                                                  \
+            i -= sizeof(TYPEW);                                               \
+            if (likely((pg >> (i & 63)) & 1)) {                               \
+                TYPEN nn = *(TYPEN *)(vn + HN(i + sizeof(TYPEN)));            \
+                *(TYPEW *)(vd + HW(i)) = OP(nn, status);                      \
+            }                                                                 \
+        } while (i & 63);                                                     \
+    } while (i != 0);                                                         \
+}
+
+DO_FCVTLT(sve2_fcvtlt_hs, uint32_t, uint16_t, H1_4, H1_2, sve_f16_to_f32)
+DO_FCVTLT(sve2_fcvtlt_sd, uint64_t, uint32_t, H1_4, H1_2, float32_to_float64)
+
+#undef DO_FCVTLT
+#undef DO_FCVTNT
target/arm/translate-sve.c +16 -0
index d410d0024e..220ff19348 100644
@@ -8121,3 +8121,19 @@ static bool trans_FCVTNT_ds(DisasContext *s, arg_rpr_esz *a)
     }
     return do_zpz_ptr(s, a->rd, a->rn, a->pg, false, gen_helper_sve2_fcvtnt_ds);
 }
+
+static bool trans_FCVTLT_hs(DisasContext *s, arg_rpr_esz *a)
+{
+    if (!dc_isar_feature(aa64_sve2, s)) {
+        return false;
+    }
+    return do_zpz_ptr(s, a->rd, a->rn, a->pg, false, gen_helper_sve2_fcvtlt_hs);
+}
+
+static bool trans_FCVTLT_sd(DisasContext *s, arg_rpr_esz *a)
+{
+    if (!dc_isar_feature(aa64_sve2, s)) {
+        return false;
+    }
+    return do_zpz_ptr(s, a->rd, a->rn, a->pg, false, gen_helper_sve2_fcvtlt_sd);
+}
Signature · 1 line
2.17.1
#3 lore raw

RFC PATCH 3/3 target/arm: Implement SVE2 FCVTXNT

In reply to Stephen Long
---
 target/arm/helper-sve.h    | 2 ++
 target/arm/sve.decode      | 1 +
 target/arm/sve_helper.c    | 7 +++++++
 target/arm/translate-sve.c | 8 ++++++++
 4 files changed, 18 insertions(+)
target/arm/helper-sve.h +2 -0
index 27a8cc2cbe..c18e4c01b0 100644
@@ -2714,6 +2714,8 @@ DEF_HELPER_FLAGS_5(sve2_sqdmlsl_idx_s, TCG_CALL_NO_RWG,
 DEF_HELPER_FLAGS_5(sve2_sqdmlsl_idx_d, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 
+DEF_HELPER_FLAGS_5(sve2_fcvtxnt, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32)
+
 DEF_HELPER_FLAGS_5(sve2_fcvtnt_sh, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 DEF_HELPER_FLAGS_5(sve2_fcvtnt_ds, TCG_CALL_NO_RWG,
target/arm/sve.decode +1 -0
index 95fc7ee32c..cb4ed70698 100644
@@ -1480,6 +1480,7 @@ STNT1_zprz      1110010 .. 10 ..... 001 ... ..... ..... \
                 @rprr_scatter_store xs=0 esz=2 scale=0
 
 ### SVE2 floating-point convert precision odd elements
+FCVTXNT         01100100 00 0010 10 101 ... ..... .....         @rd_pg_rn_e0
 FCVTNT_sh       01100100 10 0010 00 101 ... ..... .....         @rd_pg_rn_e0
 FCVTLT_hs       01100100 10 0010 01 101 ... ..... .....         @rd_pg_rn_e0
 FCVTNT_ds       01100100 11 0010 10 101 ... ..... .....         @rd_pg_rn_e0
target/arm/sve_helper.c +7 -0
index 8461d84445..56836d91cc 100644
@@ -7591,6 +7591,13 @@ void HELPER(NAME)(void *vd, void *vn, void *vg, void *status, uint32_t desc)  \
 DO_FCVTNT(sve2_fcvtnt_sh, uint32_t, uint16_t, H1_4, H1_2, sve_f32_to_f16)
 DO_FCVTNT(sve2_fcvtnt_ds, uint64_t, uint32_t, H1_4, H1_2, float64_to_float32)
 
+void HELPER(sve2_fcvtxnt)(void *vd, void *vn, void *vg,
+                          void *status, uint32_t desc)
+{
+    set_float_rounding_mode(float_round_to_odd, status);
+    HELPER(sve2_fcvtnt_ds)(vd, vn, vg, status, desc);
+}
+
 #define DO_FCVTLT(NAME, TYPEW, TYPEN, HW, HN, OP)                             \
 void HELPER(NAME)(void *vd, void *vn, void *vg, void *status, uint32_t desc)  \
 {                                                                             \
target/arm/translate-sve.c +8 -0
index 220ff19348..b1cf24f6e7 100644
@@ -8106,6 +8106,14 @@ static bool trans_SQRDCMLAH_zzzz(DisasContext *s, arg_CMLA_zzzz *a)
     return do_sve2_zzzz_fn(s, a->rd, a->rn, a->rm, a->ra, fns[a->esz], a->rot);
 }
 
+static bool trans_FCVTXNT(DisasContext *s, arg_rpr_esz *a)
+{
+    if (!dc_isar_feature(aa64_sve2, s)) {
+        return false;
+    }
+    return do_zpz_ptr(s, a->rd, a->rn, a->pg, false, gen_helper_sve2_fcvtxnt);
+}
+
 static bool trans_FCVTNT_sh(DisasContext *s, arg_rpr_esz *a)
 {
     if (!dc_isar_feature(aa64_sve2, s)) {
Signature · 1 line
2.17.1
#4 lore raw

Re: [PATCH RFC 3/3] target/arm: Implement SVE2 FCVTXNT

target/arm/sve_helper.c void HELPER(NAME)(void *vd, void *vn, void *vg, void *status, uint32_t desc) \ · line 7599

Quoted 6 lines · Stephen Long
target/arm/sve_helper.c +6 -0
+void HELPER(sve2_fcvtxnt)(void *vd, void *vn, void *vg,
+                          void *status, uint32_t desc)
+{
+    set_float_rounding_mode(float_round_to_odd, status);
+    HELPER(sve2_fcvtnt_ds)(vd, vn, vg, status, desc);
+}

This is incorrect, as it leaves the rounding mode set to odd.

I've adjusted trans_FCVTXNT to the existing do_frint_mode, which handles switching to and back from a fixed rounding mode.

r~

#5 lore raw
In reply to Stephen Long
Quoted 10 lines · Stephen Long

The only insn left in this category is BFCVTNT. I couldn't find any helper functions for bfloats so I'm not sure how to approach this insn.

Also, for insn FCVTXNT, I'm not sure if it is ok to set the status rounding mode to odd like that.

Stephen Long (3):
target/arm: Implement SVE2 FCVTNT
target/arm: Implement SVE2 FCVTLT
target/arm: Implement SVE2 FCVTXNT

Queued, with the change to the third patch as mentioned.

r~