| Message ID | 20220426125028.18844-6-lucas.araujo@eldorado.org.br |
|---|---|
| State | New |
| Headers | show |
| Series | VSX MMA Implementation | expand |
On 4/26/22 05:50, Lucas Mateus Castro(alqotel) wrote: > +#define VSXGER16(NAME, ORIG_T, OR_EL) \ > + void NAME(CPUPPCState *env, uint32_t a_r, uint32_t b_r, \ > + uint32_t at_r, uint32_t mask, uint32_t packed_flags) \ > + { \ > + ppc_vsr_t *at; \ > + float32 psum, aux_acc, va, vb, vc, vd; \ > + int i, j, xmsk_bit, ymsk_bit; \ > + uint8_t xmsk = mask & 0x0F; \ > + uint8_t ymsk = (mask >> 4) & 0x0F; \ > + uint8_t pmsk = (mask >> 8) & 0x3; \ > + ppc_vsr_t *b = cpu_vsr_ptr(env, b_r); \ > + ppc_vsr_t *a = cpu_vsr_ptr(env, a_r); \ > + float_status *excp_ptr = &env->fp_status; \ > + bool acc = ger_acc_flag(packed_flags); \ > + bool neg_acc = ger_neg_acc_flag(packed_flags); \ > + bool neg_mul = ger_neg_mul_flag(packed_flags); \ > + for (i = 0, xmsk_bit = 1 << 3; i < 4; i++, xmsk_bit >>= 1) { \ > + at = cpu_vsr_ptr(env, at_r + i); \ > + for (j = 0, ymsk_bit = 1 << 3; j < 4; j++, ymsk_bit >>= 1) {\ > + if ((xmsk_bit & xmsk) && (ymsk_bit & ymsk)) { \ > + va = !(pmsk & 2) ? float32_zero : \ > + GET_VSR(Vsr##OR_EL, a, \ > + 2 * i, ORIG_T, float32); \ > + vb = !(pmsk & 2) ? float32_zero : \ > + GET_VSR(Vsr##OR_EL, b, \ > + 2 * j, ORIG_T, float32); \ > + vc = !(pmsk & 1) ? float32_zero : \ > + GET_VSR(Vsr##OR_EL, a, \ > + 2 * i + 1, ORIG_T, float32);\ > + vd = !(pmsk & 1) ? float32_zero : \ > + GET_VSR(Vsr##OR_EL, b, \ > + 2 * j + 1, ORIG_T, float32);\ > + psum = float32_mul(va, vb, excp_ptr); \ > + psum = float32_muladd(vc, vd, psum, 0, excp_ptr); \ This isn't correct -- the intermediate 'prod' (the first multiply) is not rounded. I think the correct way to implement this (barring new softfloat functions) is to compute the intermediate product as float64 with float_round_to_odd, then float64r32_muladd into the correct rounding mode to finish. > + if (acc) { \ > + if (neg_mul) { \ > + psum = float32_neg(psum); \ > + } \ > + if (neg_acc) { \ > + aux_acc = float32_neg(at->VsrSF(j)); \ > + } else { \ > + aux_acc = at->VsrSF(j); \ > + } \ > + at->VsrSF(j) = float32_add(psum, aux_acc, \ > + excp_ptr); \ This one, thankfully, uses the rounded intermediate result 'msum', so is ok. Please do convert this from a macro. Given that float16 and bfloat16 are addressed the same, I think the only callback you need is the conversion from float16_to_float64. Drop the bf16 accessor to ppc_vsr_t. r~
On 26/04/2022 21:26, Richard Henderson wrote: > On 4/26/22 05:50, Lucas Mateus Castro(alqotel) wrote: >> +#define VSXGER16(NAME, ORIG_T, >> OR_EL) \ >> + void NAME(CPUPPCState *env, uint32_t a_r, uint32_t >> b_r, \ >> + uint32_t at_r, uint32_t mask, uint32_t >> packed_flags) \ >> + { \ >> + ppc_vsr_t >> *at; \ >> + float32 psum, aux_acc, va, vb, vc, >> vd; \ >> + int i, j, xmsk_bit, >> ymsk_bit; \ >> + uint8_t xmsk = mask & >> 0x0F; \ >> + uint8_t ymsk = (mask >> 4) & >> 0x0F; \ >> + uint8_t pmsk = (mask >> 8) & >> 0x3; \ >> + ppc_vsr_t *b = cpu_vsr_ptr(env, >> b_r); \ >> + ppc_vsr_t *a = cpu_vsr_ptr(env, >> a_r); \ >> + float_status *excp_ptr = >> &env->fp_status; \ >> + bool acc = >> ger_acc_flag(packed_flags); \ >> + bool neg_acc = >> ger_neg_acc_flag(packed_flags); \ >> + bool neg_mul = >> ger_neg_mul_flag(packed_flags); \ >> + for (i = 0, xmsk_bit = 1 << 3; i < 4; i++, xmsk_bit >>= 1) >> { \ >> + at = cpu_vsr_ptr(env, at_r + >> i); \ >> + for (j = 0, ymsk_bit = 1 << 3; j < 4; j++, ymsk_bit >>= >> 1) {\ >> + if ((xmsk_bit & xmsk) && (ymsk_bit & ymsk)) >> { \ >> + va = !(pmsk & 2) ? float32_zero >> : \ >> + GET_VSR(Vsr##OR_EL, >> a, \ >> + 2 * i, ORIG_T, >> float32); \ >> + vb = !(pmsk & 2) ? float32_zero >> : \ >> + GET_VSR(Vsr##OR_EL, >> b, \ >> + 2 * j, ORIG_T, >> float32); \ >> + vc = !(pmsk & 1) ? float32_zero >> : \ >> + GET_VSR(Vsr##OR_EL, >> a, \ >> + 2 * i + 1, ORIG_T, >> float32);\ >> + vd = !(pmsk & 1) ? float32_zero >> : \ >> + GET_VSR(Vsr##OR_EL, >> b, \ >> + 2 * j + 1, ORIG_T, >> float32);\ >> + psum = float32_mul(va, vb, >> excp_ptr); \ >> + psum = float32_muladd(vc, vd, psum, 0, >> excp_ptr); \ > > This isn't correct -- the intermediate 'prod' (the first multiply) is > not rounded. I > think the correct way to implement this (barring new softfloat > functions) is to compute > the intermediate product as float64 with float_round_to_odd, then > float64r32_muladd into > the correct rounding mode to finish. While not mentioned in the pseudocode the instruction description says: - Let prod be the single-precision product of src10 and src20 Which I understand as the result of the first multiplication being stored in a float32 But in xvbf16ger2* it's different (and I think this is the reason the last patch is resulting in the wrong signal in some 0 and inf results), the description says: - Let prod be the product of src10 and src20, having infinite precision and unbounded exponent range. - Let psum be the sum of the product, src11 multiplied by src21, and prod, having infinite precision and unbounded exponent range. - Let r1 be the value psum with its significand rounded to 24-bit precision using the rounding mode specified by RN, but retaining unbounded exponent range (i.e., cannot overflow or underflow). > >> + if (acc) >> { \ >> + if (neg_mul) >> { \ >> + psum = >> float32_neg(psum); \ >> + } \ >> + if (neg_acc) >> { \ >> + aux_acc = >> float32_neg(at->VsrSF(j)); \ >> + } else >> { \ >> + aux_acc = >> at->VsrSF(j); \ >> + } \ >> + at->VsrSF(j) = float32_add(psum, >> aux_acc, \ >> + excp_ptr); \ > > This one, thankfully, uses the rounded intermediate result 'msum', so > is ok. Yes this one is the easier one to deal with, in the description for the xvf16ger2* it specifies that msum and the result is rounded to single-precision and in the description for the xvbf16ger2 it specifies that r1 is 'rounded to a 24-bit significand precision and 8-bit exponent range (i.e., single-precision)' > > Please do convert this from a macro. Given that float16 and bfloat16 > are addressed the > same, I think the only callback you need is the conversion from > float16_to_float64. Drop > the bf16 accessor to ppc_vsr_t. > Will do, although I'm considering instead of the callback being the conversion, maybe have it be a 4 float multiplication typedef float32 mul_4float(float16, float16, float16, float16); Since float16 and bfloat16 are addressed the same, any thoughts? > > r~
On 4/27/22 14:11, Lucas Mateus Martins Araujo e Castro wrote: >> Please do convert this from a macro. Given that float16 and bfloat16 are addressed the >> same, I think the only callback you need is the conversion from float16_to_float64. Drop >> the bf16 accessor to ppc_vsr_t. >> > Will do, although I'm considering instead of the callback being the conversion, maybe have > it be a 4 float multiplication > typedef float32 mul_4float(float16, float16, float16, float16); > Since float16 and bfloat16 are addressed the same, any thoughts? The multiplication would be identical for the two types -- only the conversion is different. r~
diff --git a/include/fpu/softfloat.h b/include/fpu/softfloat.h index 3dcf20e3a2..63d7ff18f0 100644 --- a/include/fpu/softfloat.h +++ b/include/fpu/softfloat.h @@ -619,6 +619,15 @@ static inline float32 float32_chs(float32 a) return make_float32(float32_val(a) ^ 0x80000000); } +static inline float32 float32_neg(float32 a) +{ + if (((a & 0x7f800000) == 0x7f800000) && (a & 0x007fffff)) { + return a; + } else { + return float32_chs(a); + } +} + static inline bool float32_is_infinity(float32 a) { return (float32_val(a) & 0x7fffffff) == 0x7f800000; diff --git a/target/ppc/cpu.h b/target/ppc/cpu.h index b5d7b35dda..91167f8cc0 100644 --- a/target/ppc/cpu.h +++ b/target/ppc/cpu.h @@ -225,6 +225,7 @@ 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; @@ -2652,6 +2653,7 @@ 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 @@ -2663,6 +2665,7 @@ 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 6b03666d09..c3aead642a 100644 --- a/target/ppc/fpu_helper.c +++ b/target/ppc/fpu_helper.c @@ -3478,6 +3478,67 @@ static inline bool ger_neg_acc_flag(uint32_t flag) return flag & 0x4; } +#define float16_to_float32(A, PTR) float16_to_float32(A, true, PTR) + +#define GET_VSR(VSR, A, I, SRC_T, TARGET_T) \ + SRC_T##_to_##TARGET_T(A->VSR(I), excp_ptr) + +#define VSXGER16(NAME, ORIG_T, OR_EL) \ + void NAME(CPUPPCState *env, uint32_t a_r, uint32_t b_r, \ + uint32_t at_r, uint32_t mask, uint32_t packed_flags) \ + { \ + ppc_vsr_t *at; \ + float32 psum, aux_acc, va, vb, vc, vd; \ + int i, j, xmsk_bit, ymsk_bit; \ + uint8_t xmsk = mask & 0x0F; \ + uint8_t ymsk = (mask >> 4) & 0x0F; \ + uint8_t pmsk = (mask >> 8) & 0x3; \ + ppc_vsr_t *b = cpu_vsr_ptr(env, b_r); \ + ppc_vsr_t *a = cpu_vsr_ptr(env, a_r); \ + float_status *excp_ptr = &env->fp_status; \ + bool acc = ger_acc_flag(packed_flags); \ + bool neg_acc = ger_neg_acc_flag(packed_flags); \ + bool neg_mul = ger_neg_mul_flag(packed_flags); \ + for (i = 0, xmsk_bit = 1 << 3; i < 4; i++, xmsk_bit >>= 1) { \ + at = cpu_vsr_ptr(env, at_r + i); \ + for (j = 0, ymsk_bit = 1 << 3; j < 4; j++, ymsk_bit >>= 1) {\ + if ((xmsk_bit & xmsk) && (ymsk_bit & ymsk)) { \ + va = !(pmsk & 2) ? float32_zero : \ + GET_VSR(Vsr##OR_EL, a, \ + 2 * i, ORIG_T, float32); \ + vb = !(pmsk & 2) ? float32_zero : \ + GET_VSR(Vsr##OR_EL, b, \ + 2 * j, ORIG_T, float32); \ + vc = !(pmsk & 1) ? float32_zero : \ + GET_VSR(Vsr##OR_EL, a, \ + 2 * i + 1, ORIG_T, float32);\ + vd = !(pmsk & 1) ? float32_zero : \ + GET_VSR(Vsr##OR_EL, b, \ + 2 * j + 1, ORIG_T, float32);\ + psum = float32_mul(va, vb, excp_ptr); \ + psum = float32_muladd(vc, vd, psum, 0, excp_ptr); \ + if (acc) { \ + if (neg_mul) { \ + psum = float32_neg(psum); \ + } \ + if (neg_acc) { \ + aux_acc = float32_neg(at->VsrSF(j)); \ + } else { \ + aux_acc = at->VsrSF(j); \ + } \ + at->VsrSF(j) = float32_add(psum, aux_acc, \ + excp_ptr); \ + } else { \ + at->VsrSF(j) = psum; \ + } \ + } else { \ + at->VsrSF(j) = 0; \ + } \ + } \ + } \ + do_float_check_status(env, GETPC()); \ + } + #define VSXGER(NAME, TYPE, EL) \ void NAME(CPUPPCState *env, uint32_t a_r, uint32_t b_r, \ uint32_t at_r, uint32_t mask, uint32_t packed_flags) \ @@ -3522,7 +3583,11 @@ static inline bool ger_neg_acc_flag(uint32_t flag) do_float_check_status(env, GETPC()); \ } +VSXGER16(helper_XVF16GER2, float16, HF) VSXGER(helper_XVF32GER, float32, SF) VSXGER(helper_XVF64GER, float64, DF) +#undef VSXGER16 #undef VSXGER +#undef GET_VSR +#undef float16_to_float32 diff --git a/target/ppc/helper.h b/target/ppc/helper.h index 7d725292b1..cc59a3b71d 100644 --- a/target/ppc/helper.h +++ b/target/ppc/helper.h @@ -540,6 +540,7 @@ DEF_HELPER_5(XXBLENDVD, void, vsr, vsr, vsr, vsr, i32) DEF_HELPER_6(XVI4GER8, void, env, i32, i32, i32, i32, i32) DEF_HELPER_6(XVI8GER4, void, env, i32, i32, i32, i32, i32) DEF_HELPER_6(XVI16GER2, void, env, i32, i32, i32, i32, i32) +DEF_HELPER_6(XVF16GER2, void, env, i32, i32, i32, i32, i32) DEF_HELPER_6(XVF32GER, void, env, i32, i32, i32, i32, i32) DEF_HELPER_6(XVF64GER, void, env, i32, i32, i32, i32, i32) diff --git a/target/ppc/insn32.decode b/target/ppc/insn32.decode index 9652ca286c..a204730d1d 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 +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 b1fb0f31f3..9285e27159 100644 --- a/target/ppc/translate/vsx-impl.c.inc +++ b/target/ppc/translate/vsx-impl.c.inc @@ -2917,6 +2917,12 @@ TRANS64(PMXVI16GER2SPP, do_ger_MMIRR_XX3, GER_SPP, gen_helper_XVI16GER2) #define GER_PN ger_pack_flags_xvf(true, false, true) #define GER_NN ger_pack_flags_xvf(true, true, true) +TRANS(XVF16GER2, do_ger_XX3, GER_NOP, gen_helper_XVF16GER2) +TRANS(XVF16GER2PP, do_ger_XX3, GER_PP, gen_helper_XVF16GER2) +TRANS(XVF16GER2PN, do_ger_XX3, GER_PN, gen_helper_XVF16GER2) +TRANS(XVF16GER2NP, do_ger_XX3, GER_NP, gen_helper_XVF16GER2) +TRANS(XVF16GER2NN, do_ger_XX3, GER_NN, gen_helper_XVF16GER2) + TRANS(XVF32GER, do_ger_XX3, GER_NOP, gen_helper_XVF32GER) TRANS(XVF32GERPP, do_ger_XX3, GER_PP, gen_helper_XVF32GER) TRANS(XVF32GERPN, do_ger_XX3, GER_PN, gen_helper_XVF32GER)