diff mbox series

[v4,5/8] target/ppc: Implemented xvf16ger*

Message ID 20220520135129.63664-6-lucas.araujo@eldorado.org.br (mailing list archive)
State New, archived
Headers show
Series VSX MMA Implementation | expand

Commit Message

Lucas Mateus Martins Araujo e Castro May 20, 2022, 1:51 p.m. UTC 
From: "Lucas Mateus Castro (alqotel)" <lucas.araujo@eldorado.org.br>

Implement the following PowerISA v3.1 instructions:
xvf16ger2:   VSX Vector 16-bit Floating-Point GER (rank-2 update)
xvf16ger2nn: VSX Vector 16-bit Floating-Point GER (rank-2 update) Negative
multiply, Negative accumulate
xvf16ger2np: VSX Vector 16-bit Floating-Point GER (rank-2 update) Negative
multiply, Positive accumulate
xvf16ger2pn: VSX Vector 16-bit Floating-Point GER (rank-2 update) Positive
multiply, Negative accumulate
xvf16ger2pp: VSX Vector 16-bit Floating-Point GER (rank-2 update) Positive
multiply, Positive accumulate

Signed-off-by: Lucas Mateus Castro (alqotel) <lucas.araujo@eldorado.org.br>
---
It was suggested in last version of this patch that since negate is
being used to use it more, so this version tries to recreate the
operation as it's described in the ISA:
 if "[pm]xvf16ger2pp" then v2 <-bfp_ADD(r1, acc)
 if "[pm]xvf16ger2pn" then v2 <-bfp_ADD(r1, bfp_NEGATE(acc))
 if "[pm]xvf16ger2np" then v2 <-bfp_ADD(bfp_NEGATE(r1), acc)
 if "[pm]xvf16ger2nn" then v2 <-bfp_ADD(bfp_NEGATE(r1), bfp_NEGATE(acc))
A version using float32_sub instead of always add + neg was tried but in
the case of [pm]xvf16ger2np it would be necessary to invert the operations
so in a NaN - NaN situation the result could differ from the hardware.
(xvbf16ger2* behaves the same way as xvf16ger2*)
---
 target/ppc/cpu.h                    |  3 +
 target/ppc/fpu_helper.c             | 93 +++++++++++++++++++++++++++++
 target/ppc/helper.h                 |  5 ++
 target/ppc/insn32.decode            |  6 ++
 target/ppc/translate/vsx-impl.c.inc |  6 ++
 5 files changed, 113 insertions(+)

Comments

Richard Henderson May 20, 2022, 3:47 p.m. UTC | #1 
On 5/20/22 06:51, Lucas Mateus Castro(alqotel) wrote:
> +                if (acc) {
> +                    aux_acc = at[i].VsrSF(j);
> +                    if (!neg_mul && !neg_acc) {
> +                        r = float32_add(r, aux_acc, excp_ptr);
> +                    } else if (!neg_mul) {
> +                        r = float32_add(r, bfp32_neg(aux_acc), excp_ptr);
> +                    } else if (!neg_acc) {
> +                        r = float32_add(bfp32_neg(r), aux_acc, excp_ptr);
> +                    } else {
> +                        r = float32_add(bfp32_neg(r), bfp32_neg(aux_acc), excp_ptr);
> +                    }

There's no point in the 3 if's when using bfp32_neg.
Just use

   if (neg_mul) {
   }
   if (neg_acc) {
   }
   float32_add(...);

With that,
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


r~
Lucas Mateus Martins Araujo e Castro May 20, 2022, 4:42 p.m. UTC | #2 
On 20/05/2022 12:47, Richard Henderson wrote:
>
> On 5/20/22 06:51, Lucas Mateus Castro(alqotel) wrote:
>> +                if (acc) {
>> +                    aux_acc = at[i].VsrSF(j);
>> +                    if (!neg_mul && !neg_acc) {
>> +                        r = float32_add(r, aux_acc, excp_ptr);
>> +                    } else if (!neg_mul) {
>> +                        r = float32_add(r, bfp32_neg(aux_acc), 
>> excp_ptr);
>> +                    } else if (!neg_acc) {
>> +                        r = float32_add(bfp32_neg(r), aux_acc, 
>> excp_ptr);
>> +                    } else {
>> +                        r = float32_add(bfp32_neg(r), 
>> bfp32_neg(aux_acc), excp_ptr);
>> +                    }
>
> There's no point in the 3 if's when using bfp32_neg.
> Just use
>
>   if (neg_mul) {
>   }
>   if (neg_acc) {
>   }
>   float32_add(...);

You mean negate separately? Like:

     if (neg_mul) {
         r = bfp32_neg(r);
     }
     if (neg_acc) {
         aux_acc = bfp32_neg(aux_acc);
     }
     r = float32_add(r, aux_acc, excp_ptr);

If so I'll send a new version with this change later today

>
> With that,
> Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
>
>
> r~
Richard Henderson May 20, 2022, 7:03 p.m. UTC | #3 
On 5/20/22 09:42, Lucas Mateus Martins Araujo e Castro wrote:
> 
> On 20/05/2022 12:47, Richard Henderson wrote:
>>
>> On 5/20/22 06:51, Lucas Mateus Castro(alqotel) wrote:
>>> +                if (acc) {
>>> +                    aux_acc = at[i].VsrSF(j);
>>> +                    if (!neg_mul && !neg_acc) {
>>> +                        r = float32_add(r, aux_acc, excp_ptr);
>>> +                    } else if (!neg_mul) {
>>> +                        r = float32_add(r, bfp32_neg(aux_acc), excp_ptr);
>>> +                    } else if (!neg_acc) {
>>> +                        r = float32_add(bfp32_neg(r), aux_acc, excp_ptr);
>>> +                    } else {
>>> +                        r = float32_add(bfp32_neg(r), bfp32_neg(aux_acc), excp_ptr);
>>> +                    }
>>
>> There's no point in the 3 if's when using bfp32_neg.
>> Just use
>>
>>   if (neg_mul) {
>>   }
>>   if (neg_acc) {
>>   }
>>   float32_add(...);
> 
> You mean negate separately? Like:
> 
>      if (neg_mul) {
>          r = bfp32_neg(r);
>      }
>      if (neg_acc) {
>          aux_acc = bfp32_neg(aux_acc);
>      }
>      r = float32_add(r, aux_acc, excp_ptr);
> 
> If so I'll send a new version with this change later today

Yes.


r~
diff mbox series

Patch

diff --git a/target/ppc/cpu.h b/target/ppc/cpu.h
index bdedf4138e..46769a5647 100644
--- a/target/ppc/cpu.h
+++ b/target/ppc/cpu.h
@@ -227,6 +227,7 @@  typedef union _ppc_vsr_t {
     int16_t s16[8];
     int32_t s32[4];
     int64_t s64[2];
+    float16 f16[8];
     float32 f32[4];
     float64 f64[2];
     float128 f128;
@@ -2641,6 +2642,7 @@  static inline bool lsw_reg_in_range(int start, int nregs, int rx)
 #define VsrSW(i) s32[i]
 #define VsrD(i) u64[i]
 #define VsrSD(i) s64[i]
+#define VsrHF(i) f16[i]
 #define VsrSF(i) f32[i]
 #define VsrDF(i) f64[i]
 #else
@@ -2652,6 +2654,7 @@  static inline bool lsw_reg_in_range(int start, int nregs, int rx)
 #define VsrSW(i) s32[3 - (i)]
 #define VsrD(i) u64[1 - (i)]
 #define VsrSD(i) s64[1 - (i)]
+#define VsrHF(i) f16[7 - (i)]
 #define VsrSF(i) f32[3 - (i)]
 #define VsrDF(i) f64[1 - (i)]
 #endif
diff --git a/target/ppc/fpu_helper.c b/target/ppc/fpu_helper.c
index 1766da5bcf..f7da92a51a 100644
--- a/target/ppc/fpu_helper.c
+++ b/target/ppc/fpu_helper.c
@@ -36,6 +36,15 @@  static inline float128 float128_snan_to_qnan(float128 x)
 #define float32_snan_to_qnan(x) ((x) | 0x00400000)
 #define float16_snan_to_qnan(x) ((x) | 0x0200)
 
+static inline float32 bfp32_neg(float32 a)
+{
+    if (unlikely(float32_is_any_nan(a))) {
+        return a;
+    } else {
+        return float32_chs(a);
+    }
+}
+
 static inline bool fp_exceptions_enabled(CPUPPCState *env)
 {
 #ifdef CONFIG_USER_ONLY
@@ -3502,6 +3511,55 @@  static inline void vsxger_excp(CPUPPCState *env, uintptr_t retaddr)
     do_fpscr_check_status(env, retaddr);
 }
 
+typedef float64 extract_f16(float16, float_status *);
+
+static float64 extract_hf16(float16 in, float_status *fp_status)
+{
+    return float16_to_float64(in, true, fp_status);
+}
+
+static void vsxger16(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b,
+                     ppc_acc_t  *at, uint32_t mask, bool acc,
+                     bool neg_mul, bool neg_acc, extract_f16 extract)
+{
+    float32 r, aux_acc;
+    float64 psum, va, vb, vc, vd;
+    int i, j, xmsk_bit, ymsk_bit;
+    uint8_t pmsk = FIELD_EX32(mask, GER_MSK, PMSK),
+            xmsk = FIELD_EX32(mask, GER_MSK, XMSK),
+            ymsk = FIELD_EX32(mask, GER_MSK, YMSK);
+    float_status *excp_ptr = &env->fp_status;
+    for (i = 0, xmsk_bit = 1 << 3; i < 4; i++, xmsk_bit >>= 1) {
+        for (j = 0, ymsk_bit = 1 << 3; j < 4; j++, ymsk_bit >>= 1) {
+            if ((xmsk_bit & xmsk) && (ymsk_bit & ymsk)) {
+                va = !(pmsk & 2) ? float64_zero : extract(a->VsrHF(2 * i), excp_ptr);
+                vb = !(pmsk & 2) ? float64_zero : extract(b->VsrHF(2 * j), excp_ptr);
+                vc = !(pmsk & 1) ? float64_zero : extract(a->VsrHF(2 * i + 1), excp_ptr);
+                vd = !(pmsk & 1) ? float64_zero : extract(b->VsrHF(2 * j + 1), excp_ptr);
+                psum = float64_mul(va, vb, excp_ptr);
+                psum = float64r32_muladd(vc, vd, psum, 0, excp_ptr);
+                r = float64_to_float32(psum, excp_ptr);
+                if (acc) {
+                    aux_acc = at[i].VsrSF(j);
+                    if (!neg_mul && !neg_acc) {
+                        r = float32_add(r, aux_acc, excp_ptr);
+                    } else if (!neg_mul) {
+                        r = float32_add(r, bfp32_neg(aux_acc), excp_ptr);
+                    } else if (!neg_acc) {
+                        r = float32_add(bfp32_neg(r), aux_acc, excp_ptr);
+                    } else {
+                        r = float32_add(bfp32_neg(r), bfp32_neg(aux_acc), excp_ptr);
+                    }
+                }
+                at[i].VsrSF(j) = r;
+            } else {
+                at[i].VsrSF(j) = float32_zero;
+            }
+        }
+    }
+    vsxger_excp(env, GETPC());
+}
+
 typedef void vsxger_zero(ppc_vsr_t *at, int, int);
 
 typedef void vsxger_muladd_f(ppc_vsr_t *, ppc_vsr_t *, ppc_vsr_t *, int, int,
@@ -3579,6 +3637,41 @@  static void vsxger(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b, ppc_acc_t  *at,
     vsxger_excp(env, GETPC());
 }
 
+QEMU_FLATTEN
+void helper_XVF16GER2(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b,
+                     ppc_acc_t *at, uint32_t mask)
+{
+    vsxger16(env, a, b, at, mask, false, false, false, extract_hf16);
+}
+
+QEMU_FLATTEN
+void helper_XVF16GER2PP(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b,
+                        ppc_acc_t *at, uint32_t mask)
+{
+    vsxger16(env, a, b, at, mask, true, false, false, extract_hf16);
+}
+
+QEMU_FLATTEN
+void helper_XVF16GER2PN(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b,
+                        ppc_acc_t *at, uint32_t mask)
+{
+    vsxger16(env, a, b, at, mask, true, false, true, extract_hf16);
+}
+
+QEMU_FLATTEN
+void helper_XVF16GER2NP(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b,
+                        ppc_acc_t *at, uint32_t mask)
+{
+    vsxger16(env, a, b, at, mask, true, true, false, extract_hf16);
+}
+
+QEMU_FLATTEN
+void helper_XVF16GER2NN(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b,
+                        ppc_acc_t *at, uint32_t mask)
+{
+    vsxger16(env, a, b, at, mask, true, true, true, extract_hf16);
+}
+
 QEMU_FLATTEN
 void helper_XVF32GER(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b,
                      ppc_acc_t *at, uint32_t mask)
diff --git a/target/ppc/helper.h b/target/ppc/helper.h
index 054d25f3b0..7ab5ac8ee7 100644
--- a/target/ppc/helper.h
+++ b/target/ppc/helper.h
@@ -550,6 +550,11 @@  DEF_HELPER_5(XVI16GER2, void, env, vsr, vsr, acc, i32)
 DEF_HELPER_5(XVI16GER2S, void, env, vsr, vsr, acc, i32)
 DEF_HELPER_5(XVI16GER2PP, void, env, vsr, vsr, acc, i32)
 DEF_HELPER_5(XVI16GER2SPP, void, env, vsr, vsr, acc, i32)
+DEF_HELPER_5(XVF16GER2, void, env, vsr, vsr, acc, i32)
+DEF_HELPER_5(XVF16GER2PP, void, env, vsr, vsr, acc, i32)
+DEF_HELPER_5(XVF16GER2PN, void, env, vsr, vsr, acc, i32)
+DEF_HELPER_5(XVF16GER2NP, void, env, vsr, vsr, acc, i32)
+DEF_HELPER_5(XVF16GER2NN, void, env, vsr, vsr, acc, i32)
 DEF_HELPER_5(XVF32GER, void, env, vsr, vsr, acc, i32)
 DEF_HELPER_5(XVF32GERPP, void, env, vsr, vsr, acc, i32)
 DEF_HELPER_5(XVF32GERPN, void, env, vsr, vsr, acc, i32)
diff --git a/target/ppc/insn32.decode b/target/ppc/insn32.decode
index c561a17c7d..c774227d8c 100644
--- a/target/ppc/insn32.decode
+++ b/target/ppc/insn32.decode
@@ -739,6 +739,12 @@  XVI8GER4SPP     111011 ... -- ..... ..... 01100011 ..-  @XX3_at xa=%xx_xa
 XVI16GER2S      111011 ... -- ..... ..... 00101011 ..-  @XX3_at xa=%xx_xa
 XVI16GER2SPP    111011 ... -- ..... ..... 00101010 ..-  @XX3_at xa=%xx_xa
 
+XVF16GER2       111011 ... -- ..... ..... 00010011 ..-  @XX3_at xa=%xx_xa
+XVF16GER2PP     111011 ... -- ..... ..... 00010010 ..-  @XX3_at xa=%xx_xa
+XVF16GER2PN     111011 ... -- ..... ..... 10010010 ..-  @XX3_at xa=%xx_xa
+XVF16GER2NP     111011 ... -- ..... ..... 01010010 ..-  @XX3_at xa=%xx_xa
+XVF16GER2NN     111011 ... -- ..... ..... 11010010 ..-  @XX3_at xa=%xx_xa
+
 XVF32GER        111011 ... -- ..... ..... 00011011 ..-  @XX3_at xa=%xx_xa
 XVF32GERPP      111011 ... -- ..... ..... 00011010 ..-  @XX3_at xa=%xx_xa
 XVF32GERPN      111011 ... -- ..... ..... 10011010 ..-  @XX3_at xa=%xx_xa
diff --git a/target/ppc/translate/vsx-impl.c.inc b/target/ppc/translate/vsx-impl.c.inc
index 76747956bb..232a4d881e 100644
--- a/target/ppc/translate/vsx-impl.c.inc
+++ b/target/ppc/translate/vsx-impl.c.inc
@@ -2869,6 +2869,12 @@  TRANS64(PMXVI16GER2PP, do_ger, gen_helper_XVI16GER2PP)
 TRANS64(PMXVI16GER2S, do_ger, gen_helper_XVI16GER2S)
 TRANS64(PMXVI16GER2SPP, do_ger, gen_helper_XVI16GER2SPP)
 
+TRANS(XVF16GER2, do_ger, gen_helper_XVF16GER2)
+TRANS(XVF16GER2PP, do_ger, gen_helper_XVF16GER2PP)
+TRANS(XVF16GER2PN, do_ger, gen_helper_XVF16GER2PN)
+TRANS(XVF16GER2NP, do_ger, gen_helper_XVF16GER2NP)
+TRANS(XVF16GER2NN, do_ger, gen_helper_XVF16GER2NN)
+
 TRANS(XVF32GER, do_ger, gen_helper_XVF32GER)
 TRANS(XVF32GERPP, do_ger, gen_helper_XVF32GERPP)
 TRANS(XVF32GERPN, do_ger, gen_helper_XVF32GERPN)