Generate more tile sizes for QU8 gemm/igemm
- Enable 4x8 for 64 bit and 2x8 for 32 bit
- Generate QU8 2x8, 3x8, 5x8, 2x16, 3x16 and 5x16 GEMM and IGEMM microkernels
- QU8 used twice as many registers, so try different sizes that may perform better
- gemm bench and tests added new tile sizes.
PiperOrigin-RevId: 390299198
diff --git a/src/qu8-gemm/c4-neondot.c.in b/src/qu8-gemm/c4-neondot.c.in
index ce81c94..c5b14b3 100644
--- a/src/qu8-gemm/c4-neondot.c.in
+++ b/src/qu8-gemm/c4-neondot.c.in
@@ -7,7 +7,6 @@
$assert NR % 8 == 0
$assert 8 <= NR <= 16
$assert REQUANTIZATION == "RNDNU"
-$assert DATATYPE == "QU8"
#include <assert.h>
#include <arm_neon.h>
@@ -19,7 +18,7 @@
$PARAMS_STRUCT = "fp32_neonv8" if REQUANTIZATION == "FP32" else REQUANTIZATION.lower() + "_neon"
-void xnn_${DATATYPE.lower()}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
+void xnn_qu8_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_${MR}x${NR}c4__neondot(
size_t mr,
size_t nc,
size_t kc,
diff --git a/src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..6881c90
--- /dev/null
+++ b/src/qu8-gemm/gen/2x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,261 @@
+// Auto-generated file. Do not edit!
+// Template: src/qu8-gemm/c4-neondot.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot(
+ size_t mr,
+ size_t nc,
+ size_t kc,
+ const uint8_t* restrict a,
+ size_t a_stride,
+ const void* restrict w,
+ uint8_t* restrict c,
+ size_t cm_stride,
+ size_t cn_stride,
+ const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ assert(nc != 0);
+ assert(kc != 0);
+ assert(kc % sizeof(uint8_t) == 0);
+ assert(a != NULL);
+ assert(w != NULL);
+ assert(c != NULL);
+
+ kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+ const uint8_t* a0 = a;
+ uint8_t* c0 = c;
+ const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+ uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr != 2) {
+ a1 = a0;
+ c1 = c0;
+ }
+
+ const uint8x16_t vb_zero_point = vld1q_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
+
+ // Loop over groups of 16 columns.
+ do {
+ // Initialize accumulators with bias. 16 bias values are loaded from the
+ // weight matrix, at the start of the group of 16 columns.
+ uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc1x0123 = vpacc0x0123;
+ uint32x4_t vpacc1x4567 = vpacc0x4567;
+ uint32x4_t vpacc1x89AB = vpacc0x89AB;
+ uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+ uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+ uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+
+ // Inner accumulation loop along the 16 columns.
+ size_t k = kc;
+ // 2x partial unrolled loop to load 8 bytes at a time.
+ while (k >= 8 * sizeof(uint8_t)) {
+ // Load a 2x8 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+
+ // Load a 8x16 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+
+ k -= 8 * sizeof(uint8_t);
+ }
+ // Handle up to 4 final positions of `k`
+ if XNN_UNLIKELY(k != 0) {
+ // Load a 2x4 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+
+ // Load a 4x16 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+ }
+
+ // Subtract zero point accumulators with accumulators.
+ int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+ int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+ int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+ int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+ int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+ int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+ int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+ int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+
+ const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
+ const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
+ const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
+
+ vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+ vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+ vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+ vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+ vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+ vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+ vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+ vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+
+ vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+ vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+ vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+ vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+ vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+ vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+ vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+ vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+
+ vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+ vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+ vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+ vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+ vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+ vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+ vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+ vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+
+ const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+ const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+ const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+
+ uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+ uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+#else
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+ const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+ const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+
+ uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+ uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+#endif
+ const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
+ const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
+
+ vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+ vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+
+ vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+ vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+
+ if (nc >= 16) {
+ vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+ vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+
+ c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 16;
+ } else {
+ uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+ if (nc & 8) {
+ vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+ vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+ vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+ }
+ if (nc & 4) {
+ vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+ }
+ if (nc & 2) {
+ vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+ }
+ if (nc & 1) {
+ vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+ vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..e6c759f
--- /dev/null
+++ b/src/qu8-gemm/gen/2x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,193 @@
+// Auto-generated file. Do not edit!
+// Template: src/qu8-gemm/c4-neondot.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot(
+ size_t mr,
+ size_t nc,
+ size_t kc,
+ const uint8_t* restrict a,
+ size_t a_stride,
+ const void* restrict w,
+ uint8_t* restrict c,
+ size_t cm_stride,
+ size_t cn_stride,
+ const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ assert(nc != 0);
+ assert(kc != 0);
+ assert(kc % sizeof(uint8_t) == 0);
+ assert(a != NULL);
+ assert(w != NULL);
+ assert(c != NULL);
+
+ kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+ const uint8_t* a0 = a;
+ uint8_t* c0 = c;
+ const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+ uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr != 2) {
+ a1 = a0;
+ c1 = c0;
+ }
+
+ const uint8x16_t vb_zero_point = vld1q_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
+
+ // Loop over groups of 8 columns.
+ do {
+ // Initialize accumulators with bias. 8 bias values are loaded from the
+ // weight matrix, at the start of the group of 8 columns.
+ uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc1x0123 = vpacc0x0123;
+ uint32x4_t vpacc1x4567 = vpacc0x4567;
+ uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+
+ // Inner accumulation loop along the 8 columns.
+ size_t k = kc;
+ // 2x partial unrolled loop to load 8 bytes at a time.
+ while (k >= 8 * sizeof(uint8_t)) {
+ // Load a 2x8 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+
+ // Load a 8x8 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+
+ k -= 8 * sizeof(uint8_t);
+ }
+ // Handle up to 4 final positions of `k`
+ if XNN_UNLIKELY(k != 0) {
+ // Load a 2x4 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+
+ // Load a 4x8 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ }
+
+ // Subtract zero point accumulators with accumulators.
+ int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+ int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+ int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+ int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+
+ const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
+ const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
+ const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
+
+ vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+ vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+ vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+ vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+
+ vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+ vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+ vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+ vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+
+ vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+ vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+ vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+ vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+
+ const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+
+ uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+#else
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+
+ uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+#endif
+ const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
+ const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
+
+ vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+
+ vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+
+ if (nc >= 8) {
+ vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+ vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+
+ c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 8;
+ } else {
+ if (nc & 4) {
+ vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+ }
+ if (nc & 2) {
+ vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+ }
+ if (nc & 1) {
+ vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+ vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..8dafae1
--- /dev/null
+++ b/src/qu8-gemm/gen/3x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,336 @@
+// Auto-generated file. Do not edit!
+// Template: src/qu8-gemm/c4-neondot.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot(
+ size_t mr,
+ size_t nc,
+ size_t kc,
+ const uint8_t* restrict a,
+ size_t a_stride,
+ const void* restrict w,
+ uint8_t* restrict c,
+ size_t cm_stride,
+ size_t cn_stride,
+ const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ assert(nc != 0);
+ assert(kc != 0);
+ assert(kc % sizeof(uint8_t) == 0);
+ assert(a != NULL);
+ assert(w != NULL);
+ assert(c != NULL);
+
+ kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+ const uint8_t* a0 = a;
+ uint8_t* c0 = c;
+ const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+ uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 2) {
+ a1 = a0;
+ c1 = c0;
+ }
+ const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+ uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 2) {
+ a2 = a1;
+ c2 = c1;
+ }
+
+ const uint8x16_t vb_zero_point = vld1q_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
+
+ // Loop over groups of 16 columns.
+ do {
+ // Initialize accumulators with bias. 16 bias values are loaded from the
+ // weight matrix, at the start of the group of 16 columns.
+ uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc1x0123 = vpacc0x0123;
+ uint32x4_t vpacc1x4567 = vpacc0x4567;
+ uint32x4_t vpacc1x89AB = vpacc0x89AB;
+ uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+ uint32x4_t vpacc2x0123 = vpacc0x0123;
+ uint32x4_t vpacc2x4567 = vpacc0x4567;
+ uint32x4_t vpacc2x89AB = vpacc0x89AB;
+ uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
+ uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+ uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+ uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc2xCDEF = vmovq_n_u32(0);
+
+ // Inner accumulation loop along the 16 columns.
+ size_t k = kc;
+ // 2x partial unrolled loop to load 8 bytes at a time.
+ while (k >= 8 * sizeof(uint8_t)) {
+ // Load a 3x8 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+
+ // Load a 8x16 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 3x8 * 8x16 --> 3x16.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+ vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+ vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+ vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+ vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
+ vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 1);
+ vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
+ vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 1);
+
+ k -= 8 * sizeof(uint8_t);
+ }
+ // Handle up to 4 final positions of `k`
+ if XNN_UNLIKELY(k != 0) {
+ // Load a 3x4 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+
+ // Load a 4x16 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+ vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+ vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+ vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+ }
+
+ // Subtract zero point accumulators with accumulators.
+ int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+ int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+ int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+ int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+ int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+ int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+ int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+ int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+ int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+ int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+ int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x89AB));
+ int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2xCDEF));
+
+ const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
+ const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
+ const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
+
+ vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+ vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+ vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+ vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+ vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+ vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+ vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+ vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+ vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+ vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+ vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
+ vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
+
+ vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+ vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+ vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+ vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+ vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+ vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+ vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+ vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+ vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+ vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+ vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
+ vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
+
+ vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+ vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+ vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+ vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+ vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+ vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+ vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+ vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+ vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+ vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+ vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
+ vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
+
+ const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+ const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+ const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+ const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
+
+ uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+ uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+ uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
+#else
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+ const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+ const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+ const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
+
+ uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+ uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+ uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
+#endif
+ const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
+ const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
+
+ vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+ vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+ vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
+
+ vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+ vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+ vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
+
+ if (nc >= 16) {
+ vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+ vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+ vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
+
+ c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+
+ a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+ a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+
+ nc -= 16;
+ } else {
+ uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+ uint8x8_t vout2x01234567 = vget_low_u8(vout2x0123456789ABCDEF);
+ if (nc & 8) {
+ vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+ vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+ vst1_u8(c2, vout2x01234567); c2 += 8;
+ vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+ vout2x01234567 = vget_high_u8(vout2x0123456789ABCDEF);
+ }
+ if (nc & 4) {
+ vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+ vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+ vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
+ }
+ if (nc & 2) {
+ vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+ vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+ vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
+ }
+ if (nc & 1) {
+ vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+ vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+ vst1_lane_u8(c2, vout2x01234567, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..7a46793
--- /dev/null
+++ b/src/qu8-gemm/gen/3x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,239 @@
+// Auto-generated file. Do not edit!
+// Template: src/qu8-gemm/c4-neondot.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot(
+ size_t mr,
+ size_t nc,
+ size_t kc,
+ const uint8_t* restrict a,
+ size_t a_stride,
+ const void* restrict w,
+ uint8_t* restrict c,
+ size_t cm_stride,
+ size_t cn_stride,
+ const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ assert(nc != 0);
+ assert(kc != 0);
+ assert(kc % sizeof(uint8_t) == 0);
+ assert(a != NULL);
+ assert(w != NULL);
+ assert(c != NULL);
+
+ kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+ const uint8_t* a0 = a;
+ uint8_t* c0 = c;
+ const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+ uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 2) {
+ a1 = a0;
+ c1 = c0;
+ }
+ const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+ uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 2) {
+ a2 = a1;
+ c2 = c1;
+ }
+
+ const uint8x16_t vb_zero_point = vld1q_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
+
+ // Loop over groups of 8 columns.
+ do {
+ // Initialize accumulators with bias. 8 bias values are loaded from the
+ // weight matrix, at the start of the group of 8 columns.
+ uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc1x0123 = vpacc0x0123;
+ uint32x4_t vpacc1x4567 = vpacc0x4567;
+ uint32x4_t vpacc2x0123 = vpacc0x0123;
+ uint32x4_t vpacc2x4567 = vpacc0x4567;
+ uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+
+ // Inner accumulation loop along the 8 columns.
+ size_t k = kc;
+ // 2x partial unrolled loop to load 8 bytes at a time.
+ while (k >= 8 * sizeof(uint8_t)) {
+ // Load a 3x8 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+
+ // Load a 8x8 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+
+ k -= 8 * sizeof(uint8_t);
+ }
+ // Handle up to 4 final positions of `k`
+ if XNN_UNLIKELY(k != 0) {
+ // Load a 3x4 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+
+ // Load a 4x8 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ }
+
+ // Subtract zero point accumulators with accumulators.
+ int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+ int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+ int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+ int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+ int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+ int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+
+ const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
+ const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
+ const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
+
+ vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+ vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+ vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+ vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+ vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+ vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+
+ vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+ vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+ vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+ vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+ vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+ vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+
+ vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+ vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+ vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+ vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+ vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+ vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+
+ const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+
+ uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+ uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
+#else
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+
+ uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+ uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
+#endif
+ const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
+ const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
+
+ vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+ vout2x01234567 = vmax_u8(vout2x01234567, vget_low_u8(voutput_min));
+
+ vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+ vout2x01234567 = vmin_u8(vout2x01234567, vget_low_u8(voutput_max));
+
+ if (nc >= 8) {
+ vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+ vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+ vst1_u8(c2 + 0, vout2x01234567);
+
+ c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+
+ a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+ a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+
+ nc -= 8;
+ } else {
+ if (nc & 4) {
+ vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+ vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+ vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
+ }
+ if (nc & 2) {
+ vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+ vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+ vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
+ }
+ if (nc & 1) {
+ vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+ vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+ vst1_lane_u8(c2, vout2x01234567, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..7381d8a
--- /dev/null
+++ b/src/qu8-gemm/gen/5x16c4-minmax-rndnu-neondot.c
@@ -0,0 +1,482 @@
+// Auto-generated file. Do not edit!
+// Template: src/qu8-gemm/c4-neondot.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot(
+ size_t mr,
+ size_t nc,
+ size_t kc,
+ const uint8_t* restrict a,
+ size_t a_stride,
+ const void* restrict w,
+ uint8_t* restrict c,
+ size_t cm_stride,
+ size_t cn_stride,
+ const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 5);
+ assert(nc != 0);
+ assert(kc != 0);
+ assert(kc % sizeof(uint8_t) == 0);
+ assert(a != NULL);
+ assert(w != NULL);
+ assert(c != NULL);
+
+ kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+ const uint8_t* a0 = a;
+ uint8_t* c0 = c;
+ const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+ uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 2) {
+ a1 = a0;
+ c1 = c0;
+ }
+ const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+ uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 2) {
+ a2 = a1;
+ c2 = c1;
+ }
+ const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
+ uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 4) {
+ a3 = a2;
+ c3 = c2;
+ }
+ const uint8_t* a4 = (const uint8_t*) ((uintptr_t) a3 + a_stride);
+ uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 4) {
+ a4 = a3;
+ c4 = c3;
+ }
+
+ const uint8x16_t vb_zero_point = vld1q_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
+
+ // Loop over groups of 16 columns.
+ do {
+ // Initialize accumulators with bias. 16 bias values are loaded from the
+ // weight matrix, at the start of the group of 16 columns.
+ uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc1x0123 = vpacc0x0123;
+ uint32x4_t vpacc1x4567 = vpacc0x4567;
+ uint32x4_t vpacc1x89AB = vpacc0x89AB;
+ uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
+ uint32x4_t vpacc2x0123 = vpacc0x0123;
+ uint32x4_t vpacc2x4567 = vpacc0x4567;
+ uint32x4_t vpacc2x89AB = vpacc0x89AB;
+ uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
+ uint32x4_t vpacc3x0123 = vpacc0x0123;
+ uint32x4_t vpacc3x4567 = vpacc0x4567;
+ uint32x4_t vpacc3x89AB = vpacc0x89AB;
+ uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
+ uint32x4_t vpacc4x0123 = vpacc0x0123;
+ uint32x4_t vpacc4x4567 = vpacc0x4567;
+ uint32x4_t vpacc4x89AB = vpacc0x89AB;
+ uint32x4_t vpacc4xCDEF = vpacc0xCDEF;
+ uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc0xCDEF = vmovq_n_u32(0);
+ uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc1xCDEF = vmovq_n_u32(0);
+ uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc2xCDEF = vmovq_n_u32(0);
+ uint32x4_t vnacc3x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc3x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc3x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc3xCDEF = vmovq_n_u32(0);
+ uint32x4_t vnacc4x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc4x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc4x89AB = vmovq_n_u32(0);
+ uint32x4_t vnacc4xCDEF = vmovq_n_u32(0);
+
+ // Inner accumulation loop along the 16 columns.
+ size_t k = kc;
+ // 2x partial unrolled loop to load 8 bytes at a time.
+ while (k >= 8 * sizeof(uint8_t)) {
+ // Load a 5x8 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+ const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
+ const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
+
+ // Load a 8x16 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+ vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+ vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+ vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+ vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+ vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+ vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+ vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+ vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
+ vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 0);
+ vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
+ vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 0);
+ vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+ vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+ vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+ vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+ vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
+ vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 0);
+ vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
+ vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 0);
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 1);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 1);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 1);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 1);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+ vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
+ vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 1);
+ vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
+ vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 1);
+ vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
+ vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 1);
+ vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
+ vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 1);
+ vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
+ vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 1);
+ vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
+ vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 1);
+ vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
+ vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 1);
+ vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
+ vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 1);
+ vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb4567x89AB, va4x01234567, 1);
+ vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 1);
+ vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
+ vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 1);
+
+ k -= 8 * sizeof(uint8_t);
+ }
+ // Handle up to 4 final positions of `k`
+ if XNN_UNLIKELY(k != 0) {
+ // Load a 5x4 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+ const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 4;
+ const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 4;
+
+ // Load a 4x16 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
+ vnacc0x89AB = vdotq_lane_u32(vnacc0x89AB, vb_zero_point, va0x01234567, 0);
+ vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
+ vnacc0xCDEF = vdotq_lane_u32(vnacc0xCDEF, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
+ vnacc1x89AB = vdotq_lane_u32(vnacc1x89AB, vb_zero_point, va1x01234567, 0);
+ vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
+ vnacc1xCDEF = vdotq_lane_u32(vnacc1xCDEF, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
+ vnacc2x89AB = vdotq_lane_u32(vnacc2x89AB, vb_zero_point, va2x01234567, 0);
+ vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
+ vnacc2xCDEF = vdotq_lane_u32(vnacc2xCDEF, vb_zero_point, va2x01234567, 0);
+ vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+ vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+ vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+ vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+ vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
+ vnacc3x89AB = vdotq_lane_u32(vnacc3x89AB, vb_zero_point, va3x01234567, 0);
+ vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
+ vnacc3xCDEF = vdotq_lane_u32(vnacc3xCDEF, vb_zero_point, va3x01234567, 0);
+ vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+ vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+ vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+ vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+ vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
+ vnacc4x89AB = vdotq_lane_u32(vnacc4x89AB, vb_zero_point, va4x01234567, 0);
+ vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
+ vnacc4xCDEF = vdotq_lane_u32(vnacc4xCDEF, vb_zero_point, va4x01234567, 0);
+ }
+
+ // Subtract zero point accumulators with accumulators.
+ int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+ int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+ int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x89AB));
+ int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0xCDEF));
+ int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+ int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+ int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x89AB));
+ int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1xCDEF));
+ int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+ int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+ int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x89AB));
+ int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2xCDEF));
+ int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
+ int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x4567));
+ int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x89AB));
+ int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3xCDEF));
+ int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
+ int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x4567));
+ int32x4_t vacc4x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc4x89AB, vnacc4x89AB));
+ int32x4_t vacc4xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc4xCDEF, vnacc4xCDEF));
+
+ const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
+ const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
+ const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
+
+ vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+ vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+ vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
+ vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
+ vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+ vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+ vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
+ vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
+ vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+ vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+ vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
+ vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
+ vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
+ vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
+ vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
+ vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
+ vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
+ vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
+ vacc4x89AB = vshlq_s32(vacc4x89AB, vright_pre_shift);
+ vacc4xCDEF = vshlq_s32(vacc4xCDEF, vright_pre_shift);
+
+ vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+ vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+ vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
+ vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
+ vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+ vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+ vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
+ vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
+ vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+ vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+ vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
+ vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
+ vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
+ vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
+ vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
+ vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
+ vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
+ vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
+ vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
+ vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
+
+ vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+ vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+ vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
+ vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
+ vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+ vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+ vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
+ vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
+ vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+ vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+ vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
+ vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
+ vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
+ vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
+ vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
+ vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
+ vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
+ vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
+ vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
+ vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
+
+ const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+ const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+ const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+ const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
+ const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
+ const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
+ const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
+ const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
+
+ uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
+ uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
+ uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
+ uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
+ uint8x16_t vout4x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc4x89ABCDEF);
+#else
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+ const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+ const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+ const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
+ const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
+ const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
+ const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
+ const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
+
+ uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
+ uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
+ uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
+ uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
+ uint8x16_t vout4x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc4x89ABCDEF));
+#endif
+ const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
+ const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
+
+ vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
+ vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
+ vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
+ vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
+ vout4x0123456789ABCDEF = vmaxq_u8(vout4x0123456789ABCDEF, voutput_min);
+
+ vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
+ vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
+ vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
+ vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
+ vout4x0123456789ABCDEF = vminq_u8(vout4x0123456789ABCDEF, voutput_max);
+
+ if (nc >= 16) {
+ vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
+ vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
+ vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
+ vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
+ vst1q_u8(c4 + 0, vout4x0123456789ABCDEF);
+
+ c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+ c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
+ c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
+
+ a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+ a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+ a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
+ a4 = (const uint8_t*) ((uintptr_t) a4 - kc);
+
+ nc -= 16;
+ } else {
+ uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
+ uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
+ uint8x8_t vout4x01234567 = vget_low_u8(vout4x0123456789ABCDEF);
+ if (nc & 8) {
+ vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
+ vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
+ vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
+ vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
+ vst1_u8(c4, vout4x01234567); c4 += 8;
+ vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
+ vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
+ vout4x01234567 = vget_high_u8(vout4x0123456789ABCDEF);
+ }
+ if (nc & 4) {
+ vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
+ vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+ vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
+ vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
+ }
+ if (nc & 2) {
+ vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
+ vst1_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+ vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
+ vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
+ }
+ if (nc & 1) {
+ vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+ vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+ vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
+ vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
+ vst1_lane_u8(c4, vout4x01234567, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c b/src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c
new file mode 100644
index 0000000..58a9a21
--- /dev/null
+++ b/src/qu8-gemm/gen/5x8c4-minmax-rndnu-neondot.c
@@ -0,0 +1,325 @@
+// Auto-generated file. Do not edit!
+// Template: src/qu8-gemm/c4-neondot.c.in
+// Generator: tools/xngen
+//
+// Copyright 2020 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot(
+ size_t mr,
+ size_t nc,
+ size_t kc,
+ const uint8_t* restrict a,
+ size_t a_stride,
+ const void* restrict w,
+ uint8_t* restrict c,
+ size_t cm_stride,
+ size_t cn_stride,
+ const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 5);
+ assert(nc != 0);
+ assert(kc != 0);
+ assert(kc % sizeof(uint8_t) == 0);
+ assert(a != NULL);
+ assert(w != NULL);
+ assert(c != NULL);
+
+ kc = round_up_po2(kc, 4 * sizeof(uint8_t));
+ const uint8_t* a0 = a;
+ uint8_t* c0 = c;
+ const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
+ uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 2) {
+ a1 = a0;
+ c1 = c0;
+ }
+ const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
+ uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 2) {
+ a2 = a1;
+ c2 = c1;
+ }
+ const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
+ uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
+ if XNN_UNPREDICTABLE(mr < 4) {
+ a3 = a2;
+ c3 = c2;
+ }
+ const uint8_t* a4 = (const uint8_t*) ((uintptr_t) a3 + a_stride);
+ uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
+ if XNN_UNPREDICTABLE(mr <= 4) {
+ a4 = a3;
+ c4 = c3;
+ }
+
+ const uint8x16_t vb_zero_point = vld1q_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
+
+ // Loop over groups of 8 columns.
+ do {
+ // Initialize accumulators with bias. 8 bias values are loaded from the
+ // weight matrix, at the start of the group of 8 columns.
+ uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
+ uint32x4_t vpacc1x0123 = vpacc0x0123;
+ uint32x4_t vpacc1x4567 = vpacc0x4567;
+ uint32x4_t vpacc2x0123 = vpacc0x0123;
+ uint32x4_t vpacc2x4567 = vpacc0x4567;
+ uint32x4_t vpacc3x0123 = vpacc0x0123;
+ uint32x4_t vpacc3x4567 = vpacc0x4567;
+ uint32x4_t vpacc4x0123 = vpacc0x0123;
+ uint32x4_t vpacc4x4567 = vpacc0x4567;
+ uint32x4_t vnacc0x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc0x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc1x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc2x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc3x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc3x4567 = vmovq_n_u32(0);
+ uint32x4_t vnacc4x0123 = vmovq_n_u32(0);
+ uint32x4_t vnacc4x4567 = vmovq_n_u32(0);
+
+ // Inner accumulation loop along the 8 columns.
+ size_t k = kc;
+ // 2x partial unrolled loop to load 8 bytes at a time.
+ while (k >= 8 * sizeof(uint8_t)) {
+ // Load a 5x8 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
+ const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
+ const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
+
+ // Load a 8x8 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 5x8 * 8x8 --> 5x8.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+ vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+ vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+ vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+ vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+ vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+ vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+ vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 1);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 1);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 1);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 1);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 1);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 1);
+ vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
+ vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 1);
+ vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
+ vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 1);
+ vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
+ vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 1);
+ vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
+ vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 1);
+
+ k -= 8 * sizeof(uint8_t);
+ }
+ // Handle up to 4 final positions of `k`
+ if XNN_UNLIKELY(k != 0) {
+ // Load a 5x4 block of activations.
+ const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 4;
+ const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 4;
+ const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 4;
+ const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 4;
+ const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 4;
+
+ // Load a 4x8 block of weights.
+ const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+ const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
+
+ // Multiply-accumulate: 5x4 * 4x8 --> 5x8.
+ vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
+ vnacc0x0123 = vdotq_lane_u32(vnacc0x0123, vb_zero_point, va0x01234567, 0);
+ vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
+ vnacc0x4567 = vdotq_lane_u32(vnacc0x4567, vb_zero_point, va0x01234567, 0);
+ vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
+ vnacc1x0123 = vdotq_lane_u32(vnacc1x0123, vb_zero_point, va1x01234567, 0);
+ vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
+ vnacc1x4567 = vdotq_lane_u32(vnacc1x4567, vb_zero_point, va1x01234567, 0);
+ vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
+ vnacc2x0123 = vdotq_lane_u32(vnacc2x0123, vb_zero_point, va2x01234567, 0);
+ vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
+ vnacc2x4567 = vdotq_lane_u32(vnacc2x4567, vb_zero_point, va2x01234567, 0);
+ vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
+ vnacc3x0123 = vdotq_lane_u32(vnacc3x0123, vb_zero_point, va3x01234567, 0);
+ vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
+ vnacc3x4567 = vdotq_lane_u32(vnacc3x4567, vb_zero_point, va3x01234567, 0);
+ vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
+ vnacc4x0123 = vdotq_lane_u32(vnacc4x0123, vb_zero_point, va4x01234567, 0);
+ vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
+ vnacc4x4567 = vdotq_lane_u32(vnacc4x4567, vb_zero_point, va4x01234567, 0);
+ }
+
+ // Subtract zero point accumulators with accumulators.
+ int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
+ int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x4567));
+ int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
+ int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x4567));
+ int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
+ int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x4567));
+ int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
+ int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x4567));
+ int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
+ int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x4567));
+
+ const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
+ const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
+ const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
+
+ vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
+ vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
+ vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
+ vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
+ vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
+ vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
+ vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
+ vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
+ vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
+ vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
+
+ vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
+ vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
+ vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
+ vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
+ vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
+ vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
+ vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
+ vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
+ vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
+ vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
+
+ vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
+ vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
+ vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
+ vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
+ vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
+ vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
+ vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
+ vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
+ vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
+ vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
+
+ const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
+#if XNN_ARCH_ARM64
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
+ const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
+ const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
+
+ uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
+ uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
+ uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
+#else
+ const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
+ const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
+ const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
+ const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
+ const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
+
+ uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
+ uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
+ uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
+#endif
+ const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
+ const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
+
+ vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
+ vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
+ vout4x01234567 = vmax_u8(vout4x01234567, vget_low_u8(voutput_min));
+
+ vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
+ vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
+ vout4x01234567 = vmin_u8(vout4x01234567, vget_low_u8(voutput_max));
+
+ if (nc >= 8) {
+ vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
+ vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
+ vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
+ vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
+ vst1_u8(c4 + 0, vout4x01234567);
+
+ c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
+ c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
+ c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
+
+ a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
+ a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
+ a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
+ a4 = (const uint8_t*) ((uintptr_t) a4 - kc);
+
+ nc -= 8;
+ } else {
+ if (nc & 4) {
+ vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
+ vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
+ vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
+ vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
+ vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
+ }
+ if (nc & 2) {
+ vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
+ vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
+ vst1_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
+ vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
+ vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
+ vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
+ }
+ if (nc & 1) {
+ vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
+ vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
+ vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
+ vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
+ vst1_lane_u8(c4, vout4x01234567, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}