diff mbox series

[RFC,v2,7/7] target/ppc: Implemented [pm]xvbf16ger2*

Message ID 20220506121844.18969-8-lucas.araujo@eldorado.org.br
State New
Headers show
Series VSX MMA Implementation | expand

Commit Message

Lucas Mateus Martins Araujo e Castro May 6, 2022, 12:18 p.m. UTC 
From: "Lucas Mateus Castro (alqotel)" <lucas.araujo@eldorado.org.br>

Implement the following PowerISA v3.1 instructions:
xvbf16ger2:   VSX Vector bfloat16 GER (rank-2 update)
xvbf16ger2nn: VSX Vector bfloat16 GER (rank-2 update) Negative multiply,
Negative accumulate
xvbf16ger2np: VSX Vector bfloat16 GER (rank-2 update) Negative multiply,
Positive accumulate
xvbf16ger2pn: VSX Vector bfloat16 GER (rank-2 update) Positive multiply,
Negative accumulate
xvbf16ger2pp: VSX Vector bfloat16 GER (rank-2 update) Positive multiply,
Positive accumulate
pmxvbf16ger2:   Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
pmxvbf16ger2nn: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
Negative multiply, Negative accumulate
pmxvbf16ger2np: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
Negative multiply, Positive accumulate
pmxvbf16ger2pn: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
Positive multiply, Negative accumulate
pmxvbf16ger2pp: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
Positive multiply, Positive accumulate

Signed-off-by: Lucas Mateus Castro (alqotel) <lucas.araujo@eldorado.org.br>
---
There's a discrepancy between this implementation and mambo/the
hardware where implementing it with float64_mul then float64r32_muladd
sometimes results in an incorrect result after an underflow, but
implementing with float32_mul then float32_muladd results in incorrect
signal in some 0 or infinite results. I've not been able to solve this
---
 target/ppc/fpu_helper.c             | 40 +++++++++++++++++++++++++++++
 target/ppc/helper.h                 |  5 ++++
 target/ppc/insn32.decode            |  6 +++++
 target/ppc/insn64.decode            | 11 ++++++++
 target/ppc/translate/vsx-impl.c.inc | 12 +++++++++
 5 files changed, 74 insertions(+)

Comments

Richard Henderson May 8, 2022, 4:27 a.m. UTC | #1 
On 5/6/22 07:18, Lucas Mateus Castro(alqotel) wrote:
> From: "Lucas Mateus Castro (alqotel)" <lucas.araujo@eldorado.org.br>
> 
> Implement the following PowerISA v3.1 instructions:
> xvbf16ger2:   VSX Vector bfloat16 GER (rank-2 update)
> xvbf16ger2nn: VSX Vector bfloat16 GER (rank-2 update) Negative multiply,
> Negative accumulate
> xvbf16ger2np: VSX Vector bfloat16 GER (rank-2 update) Negative multiply,
> Positive accumulate
> xvbf16ger2pn: VSX Vector bfloat16 GER (rank-2 update) Positive multiply,
> Negative accumulate
> xvbf16ger2pp: VSX Vector bfloat16 GER (rank-2 update) Positive multiply,
> Positive accumulate
> pmxvbf16ger2:   Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
> pmxvbf16ger2nn: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
> Negative multiply, Negative accumulate
> pmxvbf16ger2np: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
> Negative multiply, Positive accumulate
> pmxvbf16ger2pn: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
> Positive multiply, Negative accumulate
> pmxvbf16ger2pp: Prefixed Masked VSX Vector bfloat16 GER (rank-2 update)
> Positive multiply, Positive accumulate
> 
> Signed-off-by: Lucas Mateus Castro (alqotel) <lucas.araujo@eldorado.org.br>
> ---
> There's a discrepancy between this implementation and mambo/the
> hardware where implementing it with float64_mul then float64r32_muladd
> sometimes results in an incorrect result after an underflow, but
> implementing with float32_mul then float32_muladd results in incorrect
> signal in some 0 or infinite results. I've not been able to solve this

I did suggest that the float64_mul needs to be done in round-to-odd.

Anyway, for this patch,
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


r~
Lucas Mateus Martins Araujo e Castro May 10, 2022, 5:25 p.m. UTC | #2 
On 08/05/2022 01:27, Richard Henderson wrote:
> On 5/6/22 07:18, Lucas Mateus Castro(alqotel) wrote:
>>
>> There's a discrepancy between this implementation and mambo/the
>> hardware where implementing it with float64_mul then float64r32_muladd
>> sometimes results in an incorrect result after an underflow, but
>> implementing with float32_mul then float32_muladd results in incorrect
>> signal in some 0 or infinite results. I've not been able to solve this
>
> I did suggest that the float64_mul needs to be done in round-to-odd.

 From what I understood, you meant:

     rmode = get_float_rounding_mode(&status);
     set_float_rounding_mode(float_round_to_odd, &status);
     psum = float64_mul(va, vb, &status);
     set_float_rounding_mode(rmode, &status);
     psum = float64r32_muladd(vc, vd, psum, 0, &status);

Which doesn't solve the problem, I tried other solutions but overall I 
found 3 test cases that no solution could pass all, those being:

xa = 0x 000923da 28c31f00 00018540 XXXXXXXX
xb = 0x 9d080000 000f97ac b7092f00 XXXXXXXX
xvbf16ger2 at, xa, xb
at = 0x 80000000 XXXXXXXX XXXXXXXX XXXXXXXX
         0xXXXXXXXX 80000016 XXXXXXXX XXXXXXXX
         0xXXXXXXXX XXXXXXXX 80000001 XXXXXXXX
         0xXXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX

Doing the operation either with float64 (with and without round_to_odd) 
or with a new softfloat operation that uses FloatParts64 results in 
0x80000015 instead of 0x80000016, but doing it with float32 results in 
0x00000000 instead of 0x80000000 and 0x80000002 instead of 0x80000001

Between those choices I'd go with float64 as to keep the result 
numerically close tho the actual value if the next operation treat those 
as an integer (with float32 you can end up having 0 instead of 
INT32_MIN) and the results are close if they're treated as floating-point.

>
> Anyway, for this patch,
> Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
>
>
> r~
diff mbox series

Patch

diff --git a/target/ppc/fpu_helper.c b/target/ppc/fpu_helper.c
index 6857be6ccc..0882702301 100644
--- a/target/ppc/fpu_helper.c
+++ b/target/ppc/fpu_helper.c
@@ -3491,6 +3491,11 @@  static float64 extract_hf16(float16 in, float_status *fp_status)
     return float16_to_float64(in, true, fp_status);
 }
 
+static float64 extract_bf16(bfloat16 in, float_status *fp_status)
+{
+    return bfloat16_to_float64(in, 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)
@@ -3611,6 +3616,41 @@  static void vsxger(CPUPPCState *env, ppc_vsr_t *a, ppc_vsr_t *b, ppc_acc_t  *at,
     do_float_check_status(env, GETPC());
 }
 
+QEMU_FLATTEN
+void helper_XVBF16GER2(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_bf16);
+}
+
+QEMU_FLATTEN
+void helper_XVBF16GER2PP(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_bf16);
+}
+
+QEMU_FLATTEN
+void helper_XVBF16GER2PN(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_bf16);
+}
+
+QEMU_FLATTEN
+void helper_XVBF16GER2NP(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_bf16);
+}
+
+QEMU_FLATTEN
+void helper_XVBF16GER2NN(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_bf16);
+}
+
 QEMU_FLATTEN
 void helper_XVF16GER2(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 5f2f574d30..59e6b74f94 100644
--- a/target/ppc/helper.h
+++ b/target/ppc/helper.h
@@ -551,6 +551,11 @@  DEF_HELPER_5(XVF16GER2PP, void, env, vsr, vsr, vsr, i32)
 DEF_HELPER_5(XVF16GER2PN, void, env, vsr, vsr, vsr, i32)
 DEF_HELPER_5(XVF16GER2NP, void, env, vsr, vsr, vsr, i32)
 DEF_HELPER_5(XVF16GER2NN, void, env, vsr, vsr, vsr, i32)
+DEF_HELPER_5(XVBF16GER2, void, env, vsr, vsr, vsr, i32)
+DEF_HELPER_5(XVBF16GER2PP, void, env, vsr, vsr, vsr, i32)
+DEF_HELPER_5(XVBF16GER2PN, void, env, vsr, vsr, vsr, i32)
+DEF_HELPER_5(XVBF16GER2NP, void, env, vsr, vsr, vsr, i32)
+DEF_HELPER_5(XVBF16GER2NN, void, env, vsr, vsr, vsr, i32)
 DEF_HELPER_5(XVF32GER, void, env, vsr, vsr, vsr, i32)
 DEF_HELPER_5(XVF32GERPP, void, env, vsr, vsr, vsr, i32)
 DEF_HELPER_5(XVF32GERPN, void, env, vsr, vsr, vsr, i32)
diff --git a/target/ppc/insn32.decode b/target/ppc/insn32.decode
index bbd4bc80f8..2090c17268 100644
--- a/target/ppc/insn32.decode
+++ b/target/ppc/insn32.decode
@@ -736,6 +736,12 @@  XVI8GER4SPP     111011 ... -- ..... ..... 01100011 ..-  @XX3_at xa=%xx_xa
 XVI16GER2S      111011 ... -- ..... ..... 00101011 ..-  @XX3_at xa=%xx_xa
 XVI16GER2SPP    111011 ... -- ..... ..... 00101010 ..-  @XX3_at xa=%xx_xa
 
+XVBF16GER2      111011 ... -- ..... ..... 00110011 ..-  @XX3_at xa=%xx_xa
+XVBF16GER2PP    111011 ... -- ..... ..... 00110010 ..-  @XX3_at xa=%xx_xa
+XVBF16GER2PN    111011 ... -- ..... ..... 10110010 ..-  @XX3_at xa=%xx_xa
+XVBF16GER2NP    111011 ... -- ..... ..... 01110010 ..-  @XX3_at xa=%xx_xa
+XVBF16GER2NN    111011 ... -- ..... ..... 11110010 ..-  @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
diff --git a/target/ppc/insn64.decode b/target/ppc/insn64.decode
index a12f11044c..78738924c6 100644
--- a/target/ppc/insn64.decode
+++ b/target/ppc/insn64.decode
@@ -150,6 +150,17 @@  PMXVI16GER2S    000001 11 1001 -- - - pmsk:2 ------ ........       \
 PMXVI16GER2SPP  000001 11 1001 -- - - pmsk:2 ------ ........       \
                 111011 ... -- ..... ..... 00101010 ..-  @MMIRR_XX3
 
+PMXVBF16GER2    000001 11 1001 -- - - pmsk:2 ------ ........ \
+                111011 ... -- ..... ..... 00110011 ..-  @MMIRR_XX3
+PMXVBF16GER2PP  000001 11 1001 -- - - pmsk:2 ------ ........ \
+                111011 ... -- ..... ..... 00110010 ..-  @MMIRR_XX3
+PMXVBF16GER2PN  000001 11 1001 -- - - pmsk:2 ------ ........ \
+                111011 ... -- ..... ..... 10110010 ..-  @MMIRR_XX3
+PMXVBF16GER2NP  000001 11 1001 -- - - pmsk:2 ------ ........ \
+                111011 ... -- ..... ..... 01110010 ..-  @MMIRR_XX3
+PMXVBF16GER2NN  000001 11 1001 -- - - pmsk:2 ------ ........ \
+                111011 ... -- ..... ..... 11110010 ..-  @MMIRR_XX3
+
 PMXVF16GER2     000001 11 1001 -- - - pmsk:2 ------ ........ \
                 111011 ... -- ..... ..... 00010011 ..-  @MMIRR_XX3
 PMXVF16GER2PP   000001 11 1001 -- - - pmsk:2 ------ ........ \
diff --git a/target/ppc/translate/vsx-impl.c.inc b/target/ppc/translate/vsx-impl.c.inc
index 00eed2b1b9..bb5b68ee06 100644
--- a/target/ppc/translate/vsx-impl.c.inc
+++ b/target/ppc/translate/vsx-impl.c.inc
@@ -2884,6 +2884,12 @@  TRANS64(PMXVI16GER2PP, do_ger_MMIRR_XX3, gen_helper_XVI16GER2PP)
 TRANS64(PMXVI16GER2S, do_ger_MMIRR_XX3, gen_helper_XVI16GER2S)
 TRANS64(PMXVI16GER2SPP, do_ger_MMIRR_XX3, gen_helper_XVI16GER2SPP)
 
+TRANS(XVBF16GER2, do_ger_XX3, gen_helper_XVBF16GER2)
+TRANS(XVBF16GER2PP, do_ger_XX3, gen_helper_XVBF16GER2PP)
+TRANS(XVBF16GER2PN, do_ger_XX3, gen_helper_XVBF16GER2PN)
+TRANS(XVBF16GER2NP, do_ger_XX3, gen_helper_XVBF16GER2NP)
+TRANS(XVBF16GER2NN, do_ger_XX3, gen_helper_XVBF16GER2NN)
+
 TRANS(XVF16GER2, do_ger_XX3, gen_helper_XVF16GER2)
 TRANS(XVF16GER2PP, do_ger_XX3, gen_helper_XVF16GER2PP)
 TRANS(XVF16GER2PN, do_ger_XX3, gen_helper_XVF16GER2PN)
@@ -2902,6 +2908,12 @@  TRANS(XVF64GERPN, do_ger_XX3, gen_helper_XVF64GERPN)
 TRANS(XVF64GERNP, do_ger_XX3, gen_helper_XVF64GERNP)
 TRANS(XVF64GERNN, do_ger_XX3, gen_helper_XVF64GERNN)
 
+TRANS64(PMXVBF16GER2, do_ger_MMIRR_XX3, gen_helper_XVBF16GER2)
+TRANS64(PMXVBF16GER2PP, do_ger_MMIRR_XX3, gen_helper_XVBF16GER2PP)
+TRANS64(PMXVBF16GER2PN, do_ger_MMIRR_XX3, gen_helper_XVBF16GER2PN)
+TRANS64(PMXVBF16GER2NP, do_ger_MMIRR_XX3, gen_helper_XVBF16GER2NP)
+TRANS64(PMXVBF16GER2NN, do_ger_MMIRR_XX3, gen_helper_XVBF16GER2NN)
+
 TRANS64(PMXVF16GER2, do_ger_MMIRR_XX3, gen_helper_XVF16GER2)
 TRANS64(PMXVF16GER2PP, do_ger_MMIRR_XX3, gen_helper_XVF16GER2PP)
 TRANS64(PMXVF16GER2PN, do_ger_MMIRR_XX3, gen_helper_XVF16GER2PN)