Generate QU8 GAVGPOOL microkernels from QS8 GAVGPOOL templates
PiperOrigin-RevId: 421477751
diff --git a/src/qu8-gavgpool/gen/7x-minmax-fp32-neon-c16.c b/src/qu8-gavgpool/gen/7x-minmax-fp32-neon-c16.c
new file mode 100644
index 0000000..6dac4e9
--- /dev/null
+++ b/src/qu8-gavgpool/gen/7x-minmax-fp32-neon-c16.c
@@ -0,0 +1,199 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gavgpool/unipass-neon.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/gavgpool.h>
+
+
+void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c16(
+ size_t rows,
+ size_t channels,
+ const uint8_t* input,
+ size_t input_stride,
+ const uint8_t* zero,
+ uint8_t* output,
+ const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
+{
+ assert(rows != 0);
+ assert(rows <= 7);
+ assert(channels != 0);
+
+ const uint8_t* i0 = input;
+ const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
+ if XNN_UNPREDICTABLE(rows < 2) {
+ i1 = zero;
+ }
+ const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
+ if XNN_UNPREDICTABLE(rows <= 2) {
+ i2 = zero;
+ }
+ const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
+ if XNN_UNPREDICTABLE(rows < 4) {
+ i3 = zero;
+ }
+ const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
+ if XNN_UNPREDICTABLE(rows <= 4) {
+ i4 = zero;
+ }
+ const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
+ if XNN_UNPREDICTABLE(rows < 6) {
+ i5 = zero;
+ }
+ const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
+ if XNN_UNPREDICTABLE(rows <= 6) {
+ i6 = zero;
+ }
+
+ const int32x4_t vinit_bias = vld1q_dup_s32(¶ms->fp32_neon.init_bias);
+ const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neon.scale);
+ const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
+ const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
+ const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neon.output_min);
+ const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neon.output_max);
+ for (; channels >= 16; channels -= 16) {
+ const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
+ const uint8x8_t vi0x89ABCDEF = vld1_u8(i0); i0 += 8;
+ const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
+ const uint8x8_t vi1x89ABCDEF = vld1_u8(i1); i1 += 8;
+
+ const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
+ uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
+ const uint8x8_t vi2x89ABCDEF = vld1_u8(i2); i2 += 8;
+ uint16x8_t vsum89ABCDEF = vaddl_u8(vi0x89ABCDEF, vi1x89ABCDEF);
+
+ const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
+ const uint8x8_t vi3x89ABCDEF = vld1_u8(i3); i3 += 8;
+ vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi2x89ABCDEF);
+ const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
+ const uint8x8_t vi4x89ABCDEF = vld1_u8(i4); i4 += 8;
+ vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi3x89ABCDEF);
+ const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
+ const uint8x8_t vi5x89ABCDEF = vld1_u8(i5); i5 += 8;
+ vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi4x89ABCDEF);
+ const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
+ const uint8x8_t vi6x89ABCDEF = vld1_u8(i6); i6 += 8;
+ vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi5x89ABCDEF);
+ vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
+ vsum89ABCDEF = vaddw_u8(vsum89ABCDEF, vi6x89ABCDEF);
+
+ int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
+ int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
+ int32x4_t vacc89AB = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum89ABCDEF)));
+ int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum89ABCDEF)));
+
+ float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
+ float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
+ float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
+ float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
+
+ vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
+ vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
+ vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
+ vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
+
+ vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
+ vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
+ vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
+ vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
+
+ vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
+ vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
+ vacc89AB = vqsubq_s32(vacc89AB, vmagic_bias_less_output_zero_point);
+ vaccCDEF = vqsubq_s32(vaccCDEF, vmagic_bias_less_output_zero_point);
+
+ #if XNN_ARCH_ARM64
+ int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
+ int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
+ #else // !XNN_ARCH_ARM64
+ int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
+ int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
+ #endif // !XNN_ARCH_ARM64
+
+
+ #if XNN_ARCH_ARM64
+ uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
+ #else // !XNN_ARCH_ARM64
+ uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
+ #endif // !XNN_ARCH_ARM64
+
+ vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
+
+ vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
+
+ vst1q_u8(output, vout0123456789ABCDEF); output += 16;
+ }
+ if XNN_UNLIKELY(channels != 0) {
+ do {
+ const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8;
+ const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8;
+ const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8;
+ uint16x8_t vsum01234567 = vaddl_u8(vi0x01234567, vi1x01234567);
+
+ const uint8x8_t vi3x01234567 = vld1_u8(i3); i3 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi2x01234567);
+ const uint8x8_t vi4x01234567 = vld1_u8(i4); i4 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi3x01234567);
+ const uint8x8_t vi5x01234567 = vld1_u8(i5); i5 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi4x01234567);
+ const uint8x8_t vi6x01234567 = vld1_u8(i6); i6 += 8;
+ vsum01234567 = vaddw_u8(vsum01234567, vi5x01234567);
+ vsum01234567 = vaddw_u8(vsum01234567, vi6x01234567);
+
+ int32x4_t vacc0123 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_low_u16(vsum01234567)));
+ int32x4_t vacc4567 = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget_high_u16(vsum01234567)));
+
+ float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
+ float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
+
+ vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
+ vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
+
+ vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
+ vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
+
+ vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
+ vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
+
+ #if XNN_ARCH_ARM64
+ int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
+ #else
+ int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
+ #endif
+
+ uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
+ vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
+ vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));
+
+ if XNN_LIKELY(channels >= 8) {
+ vst1_u8(output, vout01234567); output += 8;
+ channels -= 8;
+ } else {
+ if (channels & 4) {
+ vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
+ vout01234567 = vext_u8(vout01234567, vout01234567, 4);
+ }
+ if (channels & 2) {
+ vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
+ vout01234567 = vext_u8(vout01234567, vout01234567, 2);
+ }
+ if (channels & 1) {
+ vst1_lane_u8(output, vout01234567, 0); output += 1;
+ }
+ channels = 0;
+ }
+ } while (channels != 0);
+ }
+}