All of lore.kernel.org
 help / color / mirror / Atom feed
From: Richard Henderson <richard.henderson@linaro.org>
To: qemu-devel@nongnu.org
Cc: Peter Maydell <peter.maydell@linaro.org>, qemu-arm@nongnu.org
Subject: [PATCH v2 08/12] target/arm: Implement bfloat16 matrix multiply accumulate
Date: Tue, 25 May 2021 15:58:13 -0700	[thread overview]
Message-ID: <20210525225817.400336-9-richard.henderson@linaro.org> (raw)
In-Reply-To: <20210525225817.400336-1-richard.henderson@linaro.org>

This is BFMMLA for both AArch64 AdvSIMD and SVE,
and VMMLA.BF16 for AArch32 NEON.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/helper.h           |  3 +++
 target/arm/neon-shared.decode |  2 ++
 target/arm/sve.decode         |  6 +++--
 target/arm/translate-a64.c    | 10 +++++++++
 target/arm/translate-neon.c   |  9 ++++++++
 target/arm/translate-sve.c    | 12 ++++++++++
 target/arm/vec_helper.c       | 42 ++++++++++++++++++++++++++++++++++-
 7 files changed, 81 insertions(+), 3 deletions(-)

diff --git a/target/arm/helper.h b/target/arm/helper.h
index 376c1cef0f..af75d7f25f 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -1007,6 +1007,9 @@ DEF_HELPER_FLAGS_5(gvec_bfdot, TCG_CALL_NO_RWG,
 DEF_HELPER_FLAGS_5(gvec_bfdot_idx, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 
+DEF_HELPER_FLAGS_5(gvec_bfmmla, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+
 #ifdef TARGET_AARCH64
 #include "helper-a64.h"
 #include "helper-sve.h"
diff --git a/target/arm/neon-shared.decode b/target/arm/neon-shared.decode
index fa3cf14e3a..4e0a25d27c 100644
--- a/target/arm/neon-shared.decode
+++ b/target/arm/neon-shared.decode
@@ -67,6 +67,8 @@ VUMMLA         1111 1100 0.10 .... .... 1100 .1.1 .... \
                vm=%vm_dp vn=%vn_dp vd=%vd_dp
 VUSMMLA        1111 1100 1.10 .... .... 1100 .1.0 .... \
                vm=%vm_dp vn=%vn_dp vd=%vd_dp
+VMMLA_b16      1111 1100 0.00 .... .... 1100 .1.0 .... \
+               vm=%vm_dp vn=%vn_dp vd=%vd_dp
 
 VCMLA_scalar   1111 1110 0 . rot:2 .... .... 1000 . q:1 index:1 0 vm:4 \
                vn=%vn_dp vd=%vd_dp size=1
diff --git a/target/arm/sve.decode b/target/arm/sve.decode
index 51f87e8937..6c17898dee 100644
--- a/target/arm/sve.decode
+++ b/target/arm/sve.decode
@@ -1568,8 +1568,10 @@ SQRDCMLAH_zzzz  01000100 esz:2 0 rm:5 0011 rot:2 rn:5 rd:5  ra=%reg_movprfx
 USDOT_zzzz      01000100 .. 0 ..... 011 110 ..... .....  @rda_rn_rm
 
 ### SVE2 floating point matrix multiply accumulate
-
-FMMLA           01100100 .. 1 ..... 111001 ..... .....  @rda_rn_rm
+{
+  BFMMLA        01100100 01 1 ..... 111 001 ..... .....  @rda_rn_rm_e0
+  FMMLA         01100100 .. 1 ..... 111 001 ..... .....  @rda_rn_rm
+}
 
 ### SVE2 Memory Gather Load Group
 
diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
index 1df931cef5..9f8dae90ba 100644
--- a/target/arm/translate-a64.c
+++ b/target/arm/translate-a64.c
@@ -12228,6 +12228,13 @@ static void disas_simd_three_reg_same_extra(DisasContext *s, uint32_t insn)
         }
         feature = dc_isar_feature(aa64_fcma, s);
         break;
+    case 0x1d: /* BFMMLA */
+        if (size != MO_16 || !is_q) {
+            unallocated_encoding(s);
+            return;
+        }
+        feature = dc_isar_feature(aa64_bf16, s);
+        break;
     case 0x1f: /* BFDOT */
         switch (size) {
         case 1:
@@ -12321,6 +12328,9 @@ static void disas_simd_three_reg_same_extra(DisasContext *s, uint32_t insn)
         }
         return;
 
+    case 0xd: /* BFMMLA */
+        gen_gvec_op4_ool(s, is_q, rd, rn, rm, rd, 0, gen_helper_gvec_bfmmla);
+        return;
     case 0xf: /* BFDOT */
         switch (size) {
         case 1:
diff --git a/target/arm/translate-neon.c b/target/arm/translate-neon.c
index 8099767792..9d227a1e13 100644
--- a/target/arm/translate-neon.c
+++ b/target/arm/translate-neon.c
@@ -4126,3 +4126,12 @@ static bool trans_VUSMMLA(DisasContext *s, arg_VUSMMLA *a)
     return do_neon_ddda(s, 7, a->vd, a->vn, a->vm, 0,
                         gen_helper_gvec_usmmla_b);
 }
+
+static bool trans_VMMLA_b16(DisasContext *s, arg_VMMLA_b16 *a)
+{
+    if (!dc_isar_feature(aa32_bf16, s)) {
+        return false;
+    }
+    return do_neon_ddda(s, 7, a->vd, a->vn, a->vm, 0,
+                        gen_helper_gvec_bfmmla);
+}
diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c
index 6f02030635..4f575dc334 100644
--- a/target/arm/translate-sve.c
+++ b/target/arm/translate-sve.c
@@ -8677,3 +8677,15 @@ static bool trans_BFDOT_zzxz(DisasContext *s, arg_rrxr_esz *a)
     }
     return true;
 }
+
+static bool trans_BFMMLA(DisasContext *s, arg_rrrr_esz *a)
+{
+    if (!dc_isar_feature(aa64_sve_bf16, s)) {
+        return false;
+    }
+    if (sve_access_check(s)) {
+        gen_gvec_ool_zzzz(s, gen_helper_gvec_bfmmla,
+                          a->rd, a->rn, a->rm, a->ra, 0);
+    }
+    return true;
+}
diff --git a/target/arm/vec_helper.c b/target/arm/vec_helper.c
index 74a497f38c..27e9bdd329 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -2385,7 +2385,7 @@ static void do_mmla_b(void *vd, void *vn, void *vm, void *va, uint32_t desc,
          * Process the entire segment at once, writing back the
          * results only after we've consumed all of the inputs.
          *
-         * Key to indicies by column:
+         * Key to indices by column:
          *          i   j                  i             j
          */
         sum0 = a[H4(0 + 0)];
@@ -2472,3 +2472,43 @@ void HELPER(gvec_bfdot_idx)(void *vd, void *vn, void *vm,
     }
     clear_tail(d, opr_sz, simd_maxsz(desc));
 }
+
+void HELPER(gvec_bfmmla)(void *vd, void *vn, void *vm, void *va, uint32_t desc)
+{
+    intptr_t s, opr_sz = simd_oprsz(desc);
+    float32 *d = vd, *a = va;
+    uint32_t *n = vn, *m = vm;
+
+    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 indicies 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)]);
+
+        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)]);
+
+        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)]);
+
+        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)]);
+
+        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));
+}
-- 
2.25.1



  parent reply	other threads:[~2021-05-25 23:04 UTC|newest]

Thread overview: 14+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-05-25 22:58 [PATCH v2 00/12] target/arm: Implement BFloat16 Richard Henderson
2021-05-25 22:58 ` [PATCH v2 01/12] target/arm: Add isar_feature_{aa32, aa64, aa64_sve}_bf16 Richard Henderson
2021-05-25 22:58 ` [PATCH v2 02/12] target/arm: Unify unallocated path in disas_fp_1src Richard Henderson
2021-05-25 22:58 ` [PATCH v2 03/12] target/arm: Implement scalar float32 to bfloat16 conversion Richard Henderson
2021-05-25 22:58 ` [PATCH v2 04/12] target/arm: Implement vector " Richard Henderson
2021-05-25 22:58 ` [PATCH v2 05/12] softfpu: Add float_round_to_odd_inf Richard Henderson
2021-05-25 22:58 ` [PATCH v2 06/12] target/arm: Implement bfloat16 dot product (vector) Richard Henderson
2021-05-25 22:58 ` [PATCH v2 07/12] target/arm: Implement bfloat16 dot product (indexed) Richard Henderson
2021-05-25 22:58 ` Richard Henderson [this message]
2021-05-25 22:58 ` [PATCH v2 09/12] target/arm: Implement bfloat widening fma (vector) Richard Henderson
2021-05-25 22:58 ` [PATCH v2 10/12] target/arm: Implement bfloat widening fma (indexed) Richard Henderson
2021-05-25 22:58 ` [PATCH v2 11/12] linux-user/aarch64: Enable hwcap bits for bfloat16 Richard Henderson
2021-05-25 22:58 ` [PATCH v2 12/12] target/arm: Enable BFloat16 extensions Richard Henderson
2021-06-03 12:57 ` [PATCH v2 00/12] target/arm: Implement BFloat16 Peter Maydell

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210525225817.400336-9-richard.henderson@linaro.org \
    --to=richard.henderson@linaro.org \
    --cc=peter.maydell@linaro.org \
    --cc=qemu-arm@nongnu.org \
    --cc=qemu-devel@nongnu.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body. 
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.