We use bfdotadd() in four callsites for various helper functions. Currently
this all assumes that we have the FPCR.EBF=0 semantics. For FPCR.EBF=1
we will need to:
* call a different routine to bfdotadd() because we need to do a
fused multiply-add rather than separate multiply and add steps
* use a different float_status that honours the FPCR rounding mode
and denormal-flushing fields
* pass in an extra float_status that has been set up to perform
round-to-odd rounding
To prepare for this, refactor all the callsites so that instead of
for (...) {
x = bfdotadd(...);
}
they are:
float_status fpst, fpst_odd;
if (is_ebf(env, &fpst, &fpst_odd)) {
for (...) {
x = bfdotadd_ebf(..., &fpst, &fpst_odd);
}
} else {
for (...) {
x = bfdotadd(..., &fpst);
}
}
For the moment the is_ebf() function always returns false, sets up
fpst for EBF=0 semantics and never sets up fpst_odd; bfdotadd_ebf()
will assert if called. We'll fill in the handling for EBF=1 in the
next commit.
This change should be a zero-behaviour-change refactor.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
target/arm/tcg/vec_internal.h | 37 ++++++++-
target/arm/tcg/sme_helper.c | 74 ++++++++++++------
target/arm/tcg/vec_helper.c | 141 +++++++++++++++++++++++++---------
3 files changed, 192 insertions(+), 60 deletions(-)
diff --git a/target/arm/tcg/vec_internal.h b/target/arm/tcg/vec_internal.h
index 3ca1b94ccf9..094f5c169ca 100644
--- a/target/arm/tcg/vec_internal.h
+++ b/target/arm/tcg/vec_internal.h
@@ -223,13 +223,46 @@ int64_t do_sqrdmlah_d(int64_t, int64_t, int64_t, bool, bool);
* bfdotadd:
* @sum: addend
* @e1, @e2: multiplicand vectors
+ * @fpst: floating-point status to use
*
* BFloat16 2-way dot product of @e1 & @e2, accumulating with @sum.
* The @e1 and @e2 operands correspond to the 32-bit source vector
* slots and contain two Bfloat16 values each.
*
- * Corresponds to the ARM pseudocode function BFDotAdd.
+ * Corresponds to the ARM pseudocode function BFDotAdd, specialized
+ * for the FPCR.EBF == 0 case.
*/
-float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2);
+float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2, float_status *fpst);
+/**
+ * bfdotadd_ebf:
+ * @sum: addend
+ * @e1, @e2: multiplicand vectors
+ * @fpst: floating-point status to use
+ * @fpst_odd: floating-point status to use for round-to-odd operations
+ *
+ * BFloat16 2-way dot product of @e1 & @e2, accumulating with @sum.
+ * The @e1 and @e2 operands correspond to the 32-bit source vector
+ * slots and contain two Bfloat16 values each.
+ *
+ * Corresponds to the ARM pseudocode function BFDotAdd, specialized
+ * for the FPCR.EBF == 1 case.
+ */
+float32 bfdotadd_ebf(float32 sum, uint32_t e1, uint32_t e2,
+ float_status *fpst, float_status *fpst_odd);
+
+/**
+ * is_ebf:
+ * @env: CPU state
+ * @statusp: pointer to floating point status to fill in
+ * @oddstatusp: pointer to floating point status to fill in for round-to-odd
+ *
+ * Determine whether a BFDotAdd operation should use FPCR.EBF = 0
+ * or FPCR.EBF = 1 semantics. On return, has initialized *statusp
+ * and *oddstatusp to suitable float_status arguments to use with either
+ * bfdotadd() or bfdotadd_ebf().
+ * Returns true for EBF = 1, false for EBF = 0. (The caller should use this
+ * to decide whether to call bfdotadd() or bfdotadd_ebf().)
+ */
+bool is_ebf(CPUARMState *env, float_status *statusp, float_status *oddstatusp);
#endif /* TARGET_ARM_VEC_INTERNAL_H */
diff --git a/target/arm/tcg/sme_helper.c b/target/arm/tcg/sme_helper.c
index f172225b2f2..e3fbfa98fa5 100644
--- a/target/arm/tcg/sme_helper.c
+++ b/target/arm/tcg/sme_helper.c
@@ -1086,32 +1086,62 @@ void HELPER(sme_bfmopa)(CPUARMState *env, void *vza, void *vzn, void *vzm,
intptr_t row, col, oprsz = simd_maxsz(desc);
uint32_t neg = simd_data(desc) * 0x80008000u;
uint16_t *pn = vpn, *pm = vpm;
+ float_status fpst, fpst_odd;
- for (row = 0; row < oprsz; ) {
- uint16_t prow = pn[H2(row >> 4)];
- do {
- void *vza_row = vza + tile_vslice_offset(row);
- uint32_t n = *(uint32_t *)(vzn + H1_4(row));
+ if (is_ebf(env, &fpst, &fpst_odd)) {
+ for (row = 0; row < oprsz; ) {
+ uint16_t prow = pn[H2(row >> 4)];
+ do {
+ void *vza_row = vza + tile_vslice_offset(row);
+ uint32_t n = *(uint32_t *)(vzn + H1_4(row));
- n = f16mop_adj_pair(n, prow, neg);
+ n = f16mop_adj_pair(n, prow, neg);
- for (col = 0; col < oprsz; ) {
- uint16_t pcol = pm[H2(col >> 4)];
- do {
- if (prow & pcol & 0b0101) {
- uint32_t *a = vza_row + H1_4(col);
- uint32_t m = *(uint32_t *)(vzm + H1_4(col));
+ for (col = 0; col < oprsz; ) {
+ uint16_t pcol = pm[H2(col >> 4)];
+ do {
+ if (prow & pcol & 0b0101) {
+ uint32_t *a = vza_row + H1_4(col);
+ uint32_t m = *(uint32_t *)(vzm + H1_4(col));
- m = f16mop_adj_pair(m, pcol, 0);
- *a = bfdotadd(*a, n, m);
- }
- col += 4;
- pcol >>= 4;
- } while (col & 15);
- }
- row += 4;
- prow >>= 4;
- } while (row & 15);
+ m = f16mop_adj_pair(m, pcol, 0);
+ *a = bfdotadd_ebf(*a, n, m, &fpst, &fpst_odd);
+ }
+ col += 4;
+ pcol >>= 4;
+ } while (col & 15);
+ }
+ row += 4;
+ prow >>= 4;
+ } while (row & 15);
+ }
+ } else {
+ for (row = 0; row < oprsz; ) {
+ uint16_t prow = pn[H2(row >> 4)];
+ do {
+ void *vza_row = vza + tile_vslice_offset(row);
+ uint32_t n = *(uint32_t *)(vzn + H1_4(row));
+
+ n = f16mop_adj_pair(n, prow, neg);
+
+ for (col = 0; col < oprsz; ) {
+ uint16_t pcol = pm[H2(col >> 4)];
+ do {
+ if (prow & pcol & 0b0101) {
+ uint32_t *a = vza_row + H1_4(col);
+ uint32_t m = *(uint32_t *)(vzm + H1_4(col));
+
+ m = f16mop_adj_pair(m, pcol, 0);
+ *a = bfdotadd(*a, n, m, &fpst);
+ }
+ col += 4;
+ pcol >>= 4;
+ } while (col & 15);
+ }
+ row += 4;
+ prow >>= 4;
+ } while (row & 15);
+ }
}
}
diff --git a/target/arm/tcg/vec_helper.c b/target/arm/tcg/vec_helper.c
index 77efb5f47d8..baf04a0561b 100644
--- a/target/arm/tcg/vec_helper.c
+++ b/target/arm/tcg/vec_helper.c
@@ -2790,7 +2790,7 @@ DO_MMLA_B(gvec_usmmla_b, do_usmmla_b)
* BFloat16 Dot Product
*/
-float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2)
+bool is_ebf(CPUARMState *env, float_status *statusp, float_status *oddstatusp)
{
/* FPCR is ignored for BFDOT and BFMMLA. */
float_status bf_status = {
@@ -2800,29 +2800,50 @@ float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2)
.flush_inputs_to_zero = true,
.default_nan_mode = true,
};
+
+ *statusp = bf_status;
+ return false;
+}
+
+float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2, float_status *fpst)
+{
float32 t1, t2;
/*
* Extract each BFloat16 from the element pair, and shift
* them such that they become float32.
*/
- t1 = float32_mul(e1 << 16, e2 << 16, &bf_status);
- t2 = float32_mul(e1 & 0xffff0000u, e2 & 0xffff0000u, &bf_status);
- t1 = float32_add(t1, t2, &bf_status);
- t1 = float32_add(sum, t1, &bf_status);
+ t1 = float32_mul(e1 << 16, e2 << 16, fpst);
+ t2 = float32_mul(e1 & 0xffff0000u, e2 & 0xffff0000u, fpst);
+ t1 = float32_add(t1, t2, fpst);
+ t1 = float32_add(sum, t1, fpst);
return t1;
}
+float32 bfdotadd_ebf(float32 sum, uint32_t e1, uint32_t e2,
+ float_status *fpst, float_status *fpst_odd)
+{
+ g_assert_not_reached();
+}
+
void HELPER(gvec_bfdot)(void *vd, void *vn, void *vm, void *va,
void *envp, uint32_t desc)
{
+ CPUARMState *env = envp;
intptr_t i, opr_sz = simd_oprsz(desc);
float32 *d = vd, *a = va;
uint32_t *n = vn, *m = vm;
+ float_status fpst, fpst_odd;
- for (i = 0; i < opr_sz / 4; ++i) {
- d[i] = bfdotadd(a[i], n[i], m[i]);
+ if (is_ebf(env, &fpst, &fpst_odd)) {
+ for (i = 0; i < opr_sz / 4; ++i) {
+ d[i] = bfdotadd_ebf(a[i], n[i], m[i], &fpst, &fpst_odd);
+ }
+ } else {
+ for (i = 0; i < opr_sz / 4; ++i) {
+ d[i] = bfdotadd(a[i], n[i], m[i], &fpst);
+ }
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
@@ -2830,18 +2851,30 @@ void HELPER(gvec_bfdot)(void *vd, void *vn, void *vm, void *va,
void HELPER(gvec_bfdot_idx)(void *vd, void *vn, void *vm,
void *va, void *envp, uint32_t desc)
{
+ CPUARMState *env = envp;
intptr_t i, j, opr_sz = simd_oprsz(desc);
intptr_t index = simd_data(desc);
intptr_t elements = opr_sz / 4;
intptr_t eltspersegment = MIN(16 / 4, elements);
float32 *d = vd, *a = va;
uint32_t *n = vn, *m = vm;
+ float_status fpst, fpst_odd;
- for (i = 0; i < elements; i += eltspersegment) {
- uint32_t m_idx = m[i + H4(index)];
+ if (is_ebf(env, &fpst, &fpst_odd)) {
+ for (i = 0; i < elements; i += eltspersegment) {
+ uint32_t m_idx = m[i + H4(index)];
- for (j = i; j < i + eltspersegment; j++) {
- d[j] = bfdotadd(a[j], n[j], m_idx);
+ for (j = i; j < i + eltspersegment; j++) {
+ d[j] = bfdotadd_ebf(a[j], n[j], m_idx, &fpst, &fpst_odd);
+ }
+ }
+ } else {
+ for (i = 0; i < elements; i += eltspersegment) {
+ uint32_t m_idx = m[i + H4(index)];
+
+ for (j = i; j < i + eltspersegment; j++) {
+ d[j] = bfdotadd(a[j], n[j], m_idx, &fpst);
+ }
}
}
clear_tail(d, opr_sz, simd_maxsz(desc));
@@ -2850,40 +2883,76 @@ void HELPER(gvec_bfdot_idx)(void *vd, void *vn, void *vm,
void HELPER(gvec_bfmmla)(void *vd, void *vn, void *vm, void *va,
void *envp, uint32_t desc)
{
+ CPUARMState *env = envp;
intptr_t s, opr_sz = simd_oprsz(desc);
float32 *d = vd, *a = va;
uint32_t *n = vn, *m = vm;
+ float_status fpst, fpst_odd;
- for (s = 0; s < opr_sz / 4; s += 4) {
- float32 sum00, sum01, sum10, sum11;
+ if (is_ebf(env, &fpst, &fpst_odd)) {
+ for (s = 0; s < opr_sz / 4; s += 4) {
+ float32 sum00, sum01, sum10, sum11;
- /*
- * Process the entire segment at once, writing back the
- * results only after we've consumed all of the inputs.
- *
- * Key to indices by column:
- * i j i k j k
- */
- sum00 = a[s + H4(0 + 0)];
- sum00 = bfdotadd(sum00, n[s + H4(0 + 0)], m[s + H4(0 + 0)]);
- sum00 = bfdotadd(sum00, n[s + H4(0 + 1)], m[s + H4(0 + 1)]);
+ /*
+ * Process the entire segment at once, writing back the
+ * results only after we've consumed all of the inputs.
+ *
+ * Key to indices by column:
+ * i j i k j k
+ */
+ sum00 = a[s + H4(0 + 0)];
+ sum00 = bfdotadd_ebf(sum00, n[s + H4(0 + 0)], m[s + H4(0 + 0)], &fpst, &fpst_odd);
+ sum00 = bfdotadd_ebf(sum00, n[s + H4(0 + 1)], m[s + H4(0 + 1)], &fpst, &fpst_odd);
- sum01 = a[s + H4(0 + 1)];
- sum01 = bfdotadd(sum01, n[s + H4(0 + 0)], m[s + H4(2 + 0)]);
- sum01 = bfdotadd(sum01, n[s + H4(0 + 1)], m[s + H4(2 + 1)]);
+ sum01 = a[s + H4(0 + 1)];
+ sum01 = bfdotadd_ebf(sum01, n[s + H4(0 + 0)], m[s + H4(2 + 0)], &fpst, &fpst_odd);
+ sum01 = bfdotadd_ebf(sum01, n[s + H4(0 + 1)], m[s + H4(2 + 1)], &fpst, &fpst_odd);
- sum10 = a[s + H4(2 + 0)];
- sum10 = bfdotadd(sum10, n[s + H4(2 + 0)], m[s + H4(0 + 0)]);
- sum10 = bfdotadd(sum10, n[s + H4(2 + 1)], m[s + H4(0 + 1)]);
+ sum10 = a[s + H4(2 + 0)];
+ sum10 = bfdotadd_ebf(sum10, n[s + H4(2 + 0)], m[s + H4(0 + 0)], &fpst, &fpst_odd);
+ sum10 = bfdotadd_ebf(sum10, n[s + H4(2 + 1)], m[s + H4(0 + 1)], &fpst, &fpst_odd);
- sum11 = a[s + H4(2 + 1)];
- sum11 = bfdotadd(sum11, n[s + H4(2 + 0)], m[s + H4(2 + 0)]);
- sum11 = bfdotadd(sum11, n[s + H4(2 + 1)], m[s + H4(2 + 1)]);
+ sum11 = a[s + H4(2 + 1)];
+ sum11 = bfdotadd_ebf(sum11, n[s + H4(2 + 0)], m[s + H4(2 + 0)], &fpst, &fpst_odd);
+ sum11 = bfdotadd_ebf(sum11, n[s + H4(2 + 1)], m[s + H4(2 + 1)], &fpst, &fpst_odd);
- d[s + H4(0 + 0)] = sum00;
- d[s + H4(0 + 1)] = sum01;
- d[s + H4(2 + 0)] = sum10;
- d[s + H4(2 + 1)] = sum11;
+ d[s + H4(0 + 0)] = sum00;
+ d[s + H4(0 + 1)] = sum01;
+ d[s + H4(2 + 0)] = sum10;
+ d[s + H4(2 + 1)] = sum11;
+ }
+ } else {
+ for (s = 0; s < opr_sz / 4; s += 4) {
+ float32 sum00, sum01, sum10, sum11;
+
+ /*
+ * Process the entire segment at once, writing back the
+ * results only after we've consumed all of the inputs.
+ *
+ * Key to indices by column:
+ * i j i k j k
+ */
+ sum00 = a[s + H4(0 + 0)];
+ sum00 = bfdotadd(sum00, n[s + H4(0 + 0)], m[s + H4(0 + 0)], &fpst);
+ sum00 = bfdotadd(sum00, n[s + H4(0 + 1)], m[s + H4(0 + 1)], &fpst);
+
+ sum01 = a[s + H4(0 + 1)];
+ sum01 = bfdotadd(sum01, n[s + H4(0 + 0)], m[s + H4(2 + 0)], &fpst);
+ sum01 = bfdotadd(sum01, n[s + H4(0 + 1)], m[s + H4(2 + 1)], &fpst);
+
+ sum10 = a[s + H4(2 + 0)];
+ sum10 = bfdotadd(sum10, n[s + H4(2 + 0)], m[s + H4(0 + 0)], &fpst);
+ sum10 = bfdotadd(sum10, n[s + H4(2 + 1)], m[s + H4(0 + 1)], &fpst);
+
+ sum11 = a[s + H4(2 + 1)];
+ sum11 = bfdotadd(sum11, n[s + H4(2 + 0)], m[s + H4(2 + 0)], &fpst);
+ sum11 = bfdotadd(sum11, n[s + H4(2 + 1)], m[s + H4(2 + 1)], &fpst);
+
+ d[s + H4(0 + 0)] = sum00;
+ d[s + H4(0 + 1)] = sum01;
+ d[s + H4(2 + 0)] = sum10;
+ d[s + H4(2 + 1)] = sum11;
+ }
}
clear_tail(d, opr_sz, simd_maxsz(desc));
}
--
2.34.1On 7/31/24 02:03, Peter Maydell wrote:
> @@ -2790,7 +2790,7 @@ DO_MMLA_B(gvec_usmmla_b, do_usmmla_b)
> * BFloat16 Dot Product
> */
>
> -float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2)
> +bool is_ebf(CPUARMState *env, float_status *statusp, float_status *oddstatusp)
> {
> /* FPCR is ignored for BFDOT and BFMMLA. */
> float_status bf_status = {
> @@ -2800,29 +2800,50 @@ float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2)
> .flush_inputs_to_zero = true,
> .default_nan_mode = true,
> };
> +
> + *statusp = bf_status;
> + return false;
> +}
Looking at the next patch, I think dropping the local variable is better.
*statusp = (float_status){
...
};
r~
On Wed, 31 Jul 2024 at 02:48, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 7/31/24 02:03, Peter Maydell wrote:
> > @@ -2790,7 +2790,7 @@ DO_MMLA_B(gvec_usmmla_b, do_usmmla_b)
> > * BFloat16 Dot Product
> > */
> >
> > -float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2)
> > +bool is_ebf(CPUARMState *env, float_status *statusp, float_status *oddstatusp)
> > {
> > /* FPCR is ignored for BFDOT and BFMMLA. */
> > float_status bf_status = {
> > @@ -2800,29 +2800,50 @@ float32 bfdotadd(float32 sum, uint32_t e1, uint32_t e2)
> > .flush_inputs_to_zero = true,
> > .default_nan_mode = true,
> > };
> > +
> > + *statusp = bf_status;
> > + return false;
> > +}
>
> Looking at the next patch, I think dropping the local variable is better.
>
> *statusp = (float_status){
> ...
> };
Yes, I agree; I've updated this patch and the next accordingly.
-- PMM
On 7/31/24 02:03, Peter Maydell wrote:
> We use bfdotadd() in four callsites for various helper functions. Currently
> this all assumes that we have the FPCR.EBF=0 semantics. For FPCR.EBF=1
> we will need to:
> * call a different routine to bfdotadd() because we need to do a
> fused multiply-add rather than separate multiply and add steps
> * use a different float_status that honours the FPCR rounding mode
> and denormal-flushing fields
> * pass in an extra float_status that has been set up to perform
> round-to-odd rounding
>
> To prepare for this, refactor all the callsites so that instead of
> for (...) {
> x = bfdotadd(...);
> }
>
> they are:
> float_status fpst, fpst_odd;
> if (is_ebf(env, &fpst, &fpst_odd)) {
> for (...) {
> x = bfdotadd_ebf(..., &fpst, &fpst_odd);
> }
> } else {
> for (...) {
> x = bfdotadd(..., &fpst);
> }
> }
>
> For the moment the is_ebf() function always returns false, sets up
> fpst for EBF=0 semantics and never sets up fpst_odd; bfdotadd_ebf()
> will assert if called. We'll fill in the handling for EBF=1 in the
> next commit.
>
> This change should be a zero-behaviour-change refactor.
>
> Signed-off-by: Peter Maydell<peter.maydell@linaro.org>
> ---
> target/arm/tcg/vec_internal.h | 37 ++++++++-
> target/arm/tcg/sme_helper.c | 74 ++++++++++++------
> target/arm/tcg/vec_helper.c | 141 +++++++++++++++++++++++++---------
> 3 files changed, 192 insertions(+), 60 deletions(-)
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
r~
© 2016 - 2024 Red Hat, Inc.