diff mbox series

[v4,5/8] target/arm: Add helpers for FMLAL

Message ID 20190215192302.27855-6-richard.henderson@linaro.org
State New
Headers show
Series target/arm: Implement ARMv8.3-JSConv & ARMv8.2-FHM | expand

Commit Message

Richard Henderson Feb. 15, 2019, 7:22 p.m. UTC
Note that float16_to_float32 rightly squashes SNaN to QNaN.
But of course pickNaNMulAdd, for ARM, selects SNaNs first.
So we have to preserve SNaN long enough for the correct NaN
to be selected.  Thus float16_to_float32_by_bits.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

---
 target/arm/helper.h     |   5 ++
 target/arm/vec_helper.c | 114 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 119 insertions(+)

-- 
2.17.2

Comments

Peter Maydell Feb. 19, 2019, 5:31 p.m. UTC | #1
On Fri, 15 Feb 2019 at 19:23, Richard Henderson
<richard.henderson@linaro.org> wrote:
>

> Note that float16_to_float32 rightly squashes SNaN to QNaN.

> But of course pickNaNMulAdd, for ARM, selects SNaNs first.

> So we have to preserve SNaN long enough for the correct NaN

> to be selected.  Thus float16_to_float32_by_bits.

>

> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

> ---


> +/*

> + * Convert float16 to float32, raising no exceptions and

> + * preserving exceptional values, including SNaN.

> + * This is effectively an unpack+repack operation.

> + */

> +static float32 float16_to_float32_by_bits(uint32_t f16)

> +{

> +    const int f16_bias = 15;

> +    const int f32_bias = 127;

> +    uint32_t sign = extract32(f16, 15, 1);

> +    uint32_t exp = extract32(f16, 10, 5);

> +    uint32_t frac = extract32(f16, 0, 10);

> +

> +    if (exp == 0x1f) {

> +        /* Inf or NaN */

> +        exp = 0xff;

> +    } else if (exp == 0) {

> +        /* Zero or denormal.  */

> +        if (frac != 0) {

> +            /*

> +             * Denormal; these are all normal float32.

> +             * Shift the fraction so that the msb is at bit 11,

> +             * then remove bit 11 as the implicit bit of the

> +             * normalized float32.  Note that we still go through

> +             * the shift for normal numbers below, to put the

> +             * float32 fraction at the right place.

> +             */

> +            int shift = clz32(frac) - 21;

> +            frac = (frac << shift) & 0x3ff;

> +            exp = f32_bias - f16_bias - shift + 1;

> +        }

> +    } else {

> +        /* Normal number; adjust the bias.  */

> +        exp += f32_bias - f16_bias;

> +    }

> +    sign <<= 31;

> +    exp <<= 23;

> +    frac <<= 23 - 10;

> +

> +    return sign | exp | frac;

> +}


Shouldn't we be observing FPCR.FZ16 here and flushing
denormal float16 inputs to zero if it's set ?
(In the pseudocode this happens in FPUnpackBase, called
from FPUnpack.)

NB: this might be awkward because for A64 we need to use the
fpstatus with FZ16 in it (vfp.fp_status_f16) for the float16
inputs, but the one with the normal FZ bit (vfp.fp_status)
for the float32 input.

thanks
-- PMM
diff mbox series

Patch

diff --git a/target/arm/helper.h b/target/arm/helper.h
index 747cb64d29..03a613a00b 100644
--- a/target/arm/helper.h
+++ b/target/arm/helper.h
@@ -677,6 +677,11 @@  DEF_HELPER_FLAGS_5(gvec_sqsub_s, TCG_CALL_NO_RWG,
 DEF_HELPER_FLAGS_5(gvec_sqsub_d, TCG_CALL_NO_RWG,
                    void, ptr, ptr, ptr, ptr, i32)
 
+DEF_HELPER_FLAGS_5(gvec_fmlal_h, TCG_CALL_NO_RWG,
+                   void, ptr, ptr, ptr, ptr, i32)
+DEF_HELPER_FLAGS_5(gvec_fmlal_idx_h, 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/vec_helper.c b/target/arm/vec_helper.c
index dfc635cf9a..224e5315b1 100644
--- a/target/arm/vec_helper.c
+++ b/target/arm/vec_helper.c
@@ -898,3 +898,117 @@  void HELPER(gvec_sqsub_d)(void *vd, void *vq, void *vn,
     }
     clear_tail(d, oprsz, simd_maxsz(desc));
 }
+
+/*
+ * Convert float16 to float32, raising no exceptions and
+ * preserving exceptional values, including SNaN.
+ * This is effectively an unpack+repack operation.
+ */
+static float32 float16_to_float32_by_bits(uint32_t f16)
+{
+    const int f16_bias = 15;
+    const int f32_bias = 127;
+    uint32_t sign = extract32(f16, 15, 1);
+    uint32_t exp = extract32(f16, 10, 5);
+    uint32_t frac = extract32(f16, 0, 10);
+
+    if (exp == 0x1f) {
+        /* Inf or NaN */
+        exp = 0xff;
+    } else if (exp == 0) {
+        /* Zero or denormal.  */
+        if (frac != 0) {
+            /*
+             * Denormal; these are all normal float32.
+             * Shift the fraction so that the msb is at bit 11,
+             * then remove bit 11 as the implicit bit of the
+             * normalized float32.  Note that we still go through
+             * the shift for normal numbers below, to put the
+             * float32 fraction at the right place.
+             */
+            int shift = clz32(frac) - 21;
+            frac = (frac << shift) & 0x3ff;
+            exp = f32_bias - f16_bias - shift + 1;
+        }
+    } else {
+        /* Normal number; adjust the bias.  */
+        exp += f32_bias - f16_bias;
+    }
+    sign <<= 31;
+    exp <<= 23;
+    frac <<= 23 - 10;
+
+    return sign | exp | frac;
+}
+
+static uint64_t load4_f16(uint64_t *ptr, int is_q, int is_2)
+{
+    /*
+     * Branchless load of u32[0], u64[0], u32[1], or u64[1].
+     * Load the 2nd qword iff is_q & is_2.
+     * Shift to the 2nd dword iff !is_q & is_2.
+     * For !is_q & !is_2, the upper bits of the result are garbage.
+     */
+    return ptr[is_q & is_2] >> ((is_2 & ~is_q) << 5);
+}
+
+/*
+ * Note that FMLAL requires oprsz == 8 or oprsz == 16,
+ * as there is not yet SVE versions that might use blocking.
+ */
+
+void HELPER(gvec_fmlal_h)(void *vd, void *vn, void *vm,
+                          void *fpst, uint32_t desc)
+{
+    intptr_t i, oprsz = simd_oprsz(desc);
+    int is_s = extract32(desc, SIMD_DATA_SHIFT, 1);
+    int is_2 = extract32(desc, SIMD_DATA_SHIFT + 1, 1);
+    int is_q = oprsz == 16;
+    float32 *d = vd;
+    uint64_t n_4, m_4;
+
+    /* Pre-load all of the f16 data, avoiding overlap issues.  */
+    n_4 = load4_f16(vn, is_q, is_2);
+    m_4 = load4_f16(vm, is_q, is_2);
+
+    /* Negate all inputs for FMLSL at once.  */
+    if (is_s) {
+        n_4 ^= 0x8000800080008000ull;
+    }
+
+    for (i = 0; i < oprsz / 4; i++) {
+        float32 n_1 = float16_to_float32_by_bits(n_4 >> (i * 16));
+        float32 m_1 = float16_to_float32_by_bits(m_4 >> (i * 16));
+        d[H4(i)] = float32_muladd(n_1, m_1, d[H4(i)], 0, fpst);
+    }
+    clear_tail(d, oprsz, simd_maxsz(desc));
+}
+
+void HELPER(gvec_fmlal_idx_h)(void *vd, void *vn, void *vm,
+                              void *fpst, uint32_t desc)
+{
+    intptr_t i, oprsz = simd_oprsz(desc);
+    int is_s = extract32(desc, SIMD_DATA_SHIFT, 1);
+    int is_2 = extract32(desc, SIMD_DATA_SHIFT + 1, 1);
+    int index = extract32(desc, SIMD_DATA_SHIFT + 2, 3);
+    int is_q = oprsz == 16;
+    float32 *d = vd;
+    uint64_t n_4;
+    float32 m_1;
+
+    /* Pre-load all of the f16 data, avoiding overlap issues.  */
+    n_4 = load4_f16(vn, is_q, is_2);
+
+    /* Negate all inputs for FMLSL at once.  */
+    if (is_s) {
+        n_4 ^= 0x8000800080008000ull;
+    }
+
+    m_1 = float16_to_float32_by_bits(((float16 *)vm)[H2(index)]);
+
+    for (i = 0; i < oprsz / 4; i++) {
+        float32 n_1 = float16_to_float32_by_bits(n_4 >> (i * 16));
+        d[H4(i)] = float32_muladd(n_1, m_1, d[H4(i)], 0, fpst);
+    }
+    clear_tail(d, oprsz, simd_maxsz(desc));
+}