Filterbank Accumulate in ARM assembly

PiperOrigin-RevId: 470890302

7 files changed, 87 insertions, 2 deletions

diff --git a/BUILD.bazel b/BUILD.bazel
index d4ab826d9..dacf9d300 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -8310,6 +8310,7 @@ AARCH32_ASM_MICROKERNEL_SRCS = [
     "src/qu8-igemm/gen/4x8-minmax-rndnu-aarch32-neon-mlal-lane-prfm-cortex-a7.S",
     "src/qu8-igemm/gen/4x8-minmax-rndnu-aarch32-neon-mlal-lane-prfm-cortex-a53.S",
     "src/qu8-igemm/gen/4x8-minmax-rndnu-aarch32-neon-mlal-lane-prfm-ld64.S",
+    "src/u32-filterbank-accumulate/aarch32-arm-x1.S",
     "src/u32-filterbank-accumulate/aarch32-neon-x1.S",
     "src/u32-filterbank-accumulate/aarch32-neon-x2.S",
 ]
diff --git a/CMakeLists.txt b/CMakeLists.txt
index f11ac9480..9ceec5cbc 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -6752,6 +6752,7 @@ SET(AARCH32_ASM_MICROKERNEL_SRCS
   src/qu8-igemm/gen/4x8-minmax-rndnu-aarch32-neon-mlal-lane-prfm-cortex-a7.S
   src/qu8-igemm/gen/4x8-minmax-rndnu-aarch32-neon-mlal-lane-prfm-cortex-a53.S
   src/qu8-igemm/gen/4x8-minmax-rndnu-aarch32-neon-mlal-lane-prfm-ld64.S
+  src/u32-filterbank-accumulate/aarch32-arm-x1.S
   src/u32-filterbank-accumulate/aarch32-neon-x1.S
   src/u32-filterbank-accumulate/aarch32-neon-x2.S)
 
diff --git a/bench/u32-filterbank-accumulate.cc b/bench/u32-filterbank-accumulate.cc
index ba064f945..aa1133dd1 100644
--- a/bench/u32-filterbank-accumulate.cc
+++ b/bench/u32-filterbank-accumulate.cc
@@ -68,8 +68,9 @@ static void BenchmarkKernelSize(benchmark::internal::Benchmark* b)
 }
 
 #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
-BENCHMARK_CAPTURE(filterbank_accumulate, u32_aarch32_neon_x1,  xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x1,  benchmark::utils::CheckNEON)->Apply(BenchmarkKernelSize)->UseRealTime();
-BENCHMARK_CAPTURE(filterbank_accumulate, u32_aarch32_neon_x2,  xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x2,  benchmark::utils::CheckNEON)->Apply(BenchmarkKernelSize)->UseRealTime();
+BENCHMARK_CAPTURE(filterbank_accumulate, u32_aarch32_arm_x1,   xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1)->Apply(BenchmarkKernelSize)->UseRealTime();
+BENCHMARK_CAPTURE(filterbank_accumulate, u32_aarch32_neon_x1,  xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x1, benchmark::utils::CheckNEON)->Apply(BenchmarkKernelSize)->UseRealTime();
+BENCHMARK_CAPTURE(filterbank_accumulate, u32_aarch32_neon_x2,  xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x2, benchmark::utils::CheckNEON)->Apply(BenchmarkKernelSize)->UseRealTime();
 #endif  // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY
 
 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
diff --git a/src/u32-filterbank-accumulate/aarch32-arm-x1.S b/src/u32-filterbank-accumulate/aarch32-arm-x1.S
new file mode 100644
index 000000000..ba349bca0
--- /dev/null
+++ b/src/u32-filterbank-accumulate/aarch32-arm-x1.S
@@ -0,0 +1,63 @@
+// Copyright 2022 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 <xnnpack/assembly.h>
+
+.syntax unified
+
+// void xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1(
+//     size_t rows,                          r0
+//     const uint32_t* input,                r1
+//     const uint8_t* weight_widths,         r2
+//     const uint16_t* weights,              r3
+//     uint64_t* output)                     sp -> r12
+
+// d8-d15, r12-r11,r14(lr) need to be preserved if used. r13(sp),r15(pc) are reserved.
+
+// Register usage
+// input                r1  r6
+// weights              r3  r5 r7
+// weight_accumulator   r12 r8 r9
+// unweight_accumulator r10 r11
+// weight_widths        r2  r4
+
+BEGIN_FUNCTION xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1
+        .arm
+#ifndef __APPLE__
+        .arch   armv7-a
+        .fpu    neon
+#endif
+        LDR     r12, [sp]               // output
+        PUSH    {r4,r5,r6,r7,r8,r9,r10,r11}  // push 32 bytes
+        MOV     r8, 0                   // weight_accumulator
+        MOV     r9, 0
+0:
+        LDRB    r4, [r2], #1            // weight_widths
+        MOV     r10, 0                  // unweight_accumulator
+        MOV     r11, 0
+1:
+        LDR     r5, [r3], #4            // weight+unweight
+        LDR     r6, [r1], #4            // input
+        SUBS    r4, r4, #1
+        UXTH    r7, r5                  // weight
+        UXTH    r5, r5, ror #16         // unweight
+        UMLAL   r8,  r9,  r6, r7        // weight_accumulator
+        UMLAL   r10, r11, r6, r5        // unweight_accumulator
+        BHI     1b
+
+        STMIA   r12!, {r8, r9}
+        SUBS    r0, r0, #1
+        MOV     r8, r10                 // weight_accumulator = unweight_accumulator
+        MOV     r9, r11
+        BNE     0b
+
+        POP     {r4,r5,r6,r7,r8,r9,r10,r11}
+        BX      lr
+
+END_FUNCTION xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1
+
+#ifdef __ELF__
+.section ".note.GNU-stack","",%progbits
+#endif
diff --git a/src/xnnpack/filterbank.h b/src/xnnpack/filterbank.h
index f6d7a59ef..dae11b92a 100644
--- a/src/xnnpack/filterbank.h
+++ b/src/xnnpack/filterbank.h
@@ -23,6 +23,7 @@ extern "C" {
     const uint16_t* weights,                                        \
     uint64_t* output);
 
+DECLARE_U32_FILTERBANK_ACCUMULATE_UKERNEL_FUNCTION(xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1)
 DECLARE_U32_FILTERBANK_ACCUMULATE_UKERNEL_FUNCTION(xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x1)
 DECLARE_U32_FILTERBANK_ACCUMULATE_UKERNEL_FUNCTION(xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x2)
 DECLARE_U32_FILTERBANK_ACCUMULATE_UKERNEL_FUNCTION(xnn_u32_filterbank_accumulate_ukernel__neon_x1)
diff --git a/test/u32-filterbank-accumulate.cc b/test/u32-filterbank-accumulate.cc
index 0669af136..50bfca124 100644
--- a/test/u32-filterbank-accumulate.cc
+++ b/test/u32-filterbank-accumulate.cc
@@ -18,6 +18,23 @@
 
 
 #if XNN_ARCH_ARM
+  TEST(U32_FILTERBANK_ACCUMULATE__AARCH32_ARM_X1, rows_eq_1) {
+    FilterbankAccumulateMicrokernelTester()
+      .rows(1)
+      .Test(xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1);
+  }
+
+  TEST(U32_FILTERBANK_ACCUMULATE__AARCH32_ARM_X1, rows_gt_1) {
+    for (size_t rows = 2; rows < 10; rows++) {
+      FilterbankAccumulateMicrokernelTester()
+        .rows(rows)
+        .Test(xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1);
+    }
+  }
+#endif  // XNN_ARCH_ARM
+
+
+#if XNN_ARCH_ARM
   TEST(U32_FILTERBANK_ACCUMULATE__AARCH32_NEON_X1, rows_eq_1) {
     TEST_REQUIRES_ARM_NEON;
     FilterbankAccumulateMicrokernelTester()
diff --git a/test/u32-filterbank-accumulate.yaml b/test/u32-filterbank-accumulate.yaml
index 54798504c..332cbd350 100644
--- a/test/u32-filterbank-accumulate.yaml
+++ b/test/u32-filterbank-accumulate.yaml
@@ -5,6 +5,7 @@
 
 
 # AArch32 assembly
+- name: xnn_u32_filterbank_accumulate_ukernel__aarch32_arm_x1
 - name: xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x1
 - name: xnn_u32_filterbank_accumulate_ukernel__aarch32_neon_x2