aboutsummaryrefslogtreecommitdiff
path: root/arm_compute/core/NEON/NEAsymm.h
diff options
context:
space:
mode:
authorGeorgios Pinitas <georgios.pinitas@arm.com>2019-11-11 18:24:22 +0000
committerGeorgios Pinitas <georgios.pinitas@arm.com>2019-11-12 21:22:47 +0000
commit8217c8e4f488eb32733c481ab3a4d905069479f1 (patch)
treef54e10c459d5bc5b847ef0caba075a636d9f1df5 /arm_compute/core/NEON/NEAsymm.h
parent94e0cf960ea6116eb57fa88d9b951f859b52c602 (diff)
downloadComputeLibrary-8217c8e4f488eb32733c481ab3a4d905069479f1.tar.gz
COMPMID-2895: Remove QASYMM8_PER_CHANNEL data type
Change-Id: I2d1b77370f8eceeaeae95306b4db5d90ababb76f Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com> Reviewed-on: https://review.mlplatform.org/c/2266 Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'arm_compute/core/NEON/NEAsymm.h')
-rw-r--r--arm_compute/core/NEON/NEAsymm.h17
1 files changed, 8 insertions, 9 deletions
diff --git a/arm_compute/core/NEON/NEAsymm.h b/arm_compute/core/NEON/NEAsymm.h
index a3bd7e28f0..c75a58046b 100644
--- a/arm_compute/core/NEON/NEAsymm.h
+++ b/arm_compute/core/NEON/NEAsymm.h
@@ -325,23 +325,22 @@ inline float32x4x4_t vdequantize(const uint8x16_t &qv, float scale, int32_t offs
return vdequantized_input;
}
-/** Dequantize following an asymmetric quantization scheme a neon vector holding 16 quantized values.
+/** Dequantize following symmetric quantization scheme a neon vector holding 16 quantized values.
*
- * @param[in] qv Input values to be dequantized.
- * @param[in] vscale Vector containing quantization scaling factors.
- * @param[in] voffset Vector containing quantization offset.
+ * @param[in] qv Input values to be dequantized.
+ * @param[in] vscale Vector containing quantization scaling factors.
*
* @return Dequantized values in a neon vector
*/
-inline float32x4x4_t vdequantize(const uint8x16_t &qv, const float32x4x4_t vscale, const int32x4x4_t voffset)
+inline float32x4x4_t vdequantize(const int8x16_t &qv, const float32x4x4_t vscale)
{
const float32x4x4_t vdequantized_input =
{
{
- vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(qv))))), voffset.val[0])), vscale.val[0]),
- vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(qv))))), voffset.val[1])), vscale.val[1]),
- vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(qv))))), voffset.val[2])), vscale.val[2]),
- vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(qv))))), voffset.val[3])), vscale.val[3]),
+ vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(qv))))), vscale.val[0]),
+ vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(qv))))), vscale.val[1]),
+ vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(qv))))), vscale.val[2]),
+ vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(qv))))), vscale.val[3]),
}
};
return vdequantized_input;