Add read bandwidth and small refactor.

- Addition of read bandwidth test.
- Move the allocation of all buffers used for testing into the
  setSize() function.
- Add a canRun function to indicate whether a test is able to
  run on this platform. Currently, this only disables neon bandwidth
  tests on platforms that don't support neon.
- Refactor the thread testing functions to use one common routine.

Merge from internal master.

(cherry-picked from 1a3794a84074d7f22b8ddaba840aedd758a14cdd)

Change-Id: Ie0f29ec7e484a376471d121333b1982be8f41393

3 files changed, 359 insertions, 102 deletions

diff --git a/tests/memtest/bandwidth.cpp b/tests/memtest/bandwidth.cpp
index d0f09108..0e0ec2e0 100644
--- a/tests/memtest/bandwidth.cpp
+++ b/tests/memtest/bandwidth.cpp
@@ -98,34 +98,48 @@ BandwidthBenchmark *createBandwidthBenchmarkObject(arg_t values) {
         size = values["size"].int_value;
     }
     if (strcmp(name, "copy_ldrd_strd") == 0) {
-        bench = new CopyLdrdStrdBenchmark(size);
+        bench = new CopyLdrdStrdBenchmark();
     } else if (strcmp(name, "copy_ldmia_stmia") == 0) {
-        bench = new CopyLdmiaStmiaBenchmark(size);
+        bench = new CopyLdmiaStmiaBenchmark();
     } else if (strcmp(name, "copy_vld_vst") == 0) {
-        bench = new CopyVldVstBenchmark(size);
+        bench = new CopyVldVstBenchmark();
     } else if (strcmp(name, "copy_vldmia_vstmia") == 0) {
-        bench = new CopyVldmiaVstmiaBenchmark(size);
+        bench = new CopyVldmiaVstmiaBenchmark();
     } else if (strcmp(name, "memcpy") == 0) {
-        bench = new MemcpyBenchmark(size);
+        bench = new MemcpyBenchmark();
     } else if (strcmp(name, "write_strd") == 0) {
-        bench = new WriteStrdBenchmark(size);
+        bench = new WriteStrdBenchmark();
     } else if (strcmp(name, "write_stmia") == 0) {
-        bench = new WriteStmiaBenchmark(size);
+        bench = new WriteStmiaBenchmark();
     } else if (strcmp(name, "write_vst") == 0) {
-        bench = new WriteVstBenchmark(size);
+        bench = new WriteVstBenchmark();
     } else if (strcmp(name, "write_vstmia") == 0) {
-        bench = new WriteVstmiaBenchmark(size);
+        bench = new WriteVstmiaBenchmark();
     } else if (strcmp(name, "memset") == 0) {
-        bench = new MemsetBenchmark(size);
+        bench = new MemsetBenchmark();
+    } else if (strcmp(name, "read_ldrd") == 0) {
+        bench = new ReadLdrdBenchmark();
+    } else if (strcmp(name, "read_ldmia") == 0) {
+        bench = new ReadLdmiaBenchmark();
+    } else if (strcmp(name, "read_vld") == 0) {
+        bench = new ReadVldBenchmark();
+    } else if (strcmp(name, "read_vldmia") == 0) {
+        bench = new ReadVldmiaBenchmark();
+    } else {
+        printf("Unknown type name %s\n", name);
+        return NULL;
     }
 
-    if (bench) {
-        if (values.count("num_warm_loops") > 0) {
-            bench->set_num_loops(values["num_warm_loops"].int_value);
-        }
-        if (values.count("num_loops") > 0) {
-            bench->set_num_loops(values["num_loops"].int_value);
-        }
+    if (!bench->setSize(values["size"].int_value)) {
+        printf("Failed to allocate buffers for benchmark.\n");
+        return NULL;
+    }
+
+    if (values.count("num_warm_loops") > 0) {
+        bench->set_num_loops(values["num_warm_loops"].int_value);
+    }
+    if (values.count("num_loops") > 0) {
+        bench->set_num_loops(values["num_loops"].int_value);
     }
 
     return bench;
@@ -208,7 +222,6 @@ bool processThreadArgs(int argc, char** argv, option_t options[],
 
     BandwidthBenchmark *bench = createBandwidthBenchmarkObject(*values);
     if (!bench) {
-        printf("Unknown type %s\n", (*values)["type"].char_value);
         return false;
     }
 
@@ -292,6 +305,9 @@ int per_core_bandwidth(int argc, char** argv) {
          it != cpu_list.end(); ++it, ++i) {
         args[i].core = *it;
         args[i].bench = createBandwidthBenchmarkObject(values);
+        if (!args[i].bench) {
+            return 0;
+        }
     }
 
     printf("Running on %d cores\n", cpu_list.size());
@@ -325,6 +341,9 @@ int multithread_bandwidth(int argc, char** argv) {
     for (int i = 0; i < num_threads; i++) {
         args[i].core = -1;
         args[i].bench = createBandwidthBenchmarkObject(values);
+        if (!args[i].bench) {
+            return 0;
+        }
     }
 
     printf("Running %d threads\n", num_threads);
@@ -341,18 +360,20 @@ int multithread_bandwidth(int argc, char** argv) {
     return 0;
 }
 
-int copy_bandwidth(int argc, char** argv) {
+bool run_bandwidth_benchmark(int argc, char** argv, const char *name,
+                             std::vector<BandwidthBenchmark*> bench_objs) {
     arg_t values;
     values["size"].int_value = 0;
-    values["num_loops"].int_value = BandwidthBenchmark::DEFAULT_NUM_LOOPS;
-    values["num_warm_loops"].int_value = BandwidthBenchmark::DEFAULT_NUM_WARM_LOOPS;
+    values["num_warm_loops"].int_value = 0;
+    values["num_loops"].int_value = 0;
     if (!processBandwidthOptions(argc, argv, bandwidth_opts, &values)) {
         return -1;
     }
+
     size_t size = values["size"].int_value;
     if ((size % 64) != 0) {
         printf("The size value must be a multiple of 64.\n");
-        return -1;
+        return 1;
     }
 
     if (setpriority(PRIO_PROCESS, 0, -20)) {
@@ -360,68 +381,77 @@ int copy_bandwidth(int argc, char** argv) {
         return -1;
     }
 
-    std::vector<BandwidthBenchmark*> bench_objs;
-    bench_objs.push_back(new CopyLdrdStrdBenchmark(size));
-    bench_objs.push_back(new CopyLdmiaStmiaBenchmark(size));
-    bench_objs.push_back(new CopyVldVstBenchmark(size));
-    bench_objs.push_back(new CopyVldmiaVstmiaBenchmark(size));
-    bench_objs.push_back(new MemcpyBenchmark(size));
-
-    printf("Benchmarking copy bandwidth\n");
-    printf("  size = %d\n", bench_objs[0]->size());
-    printf("  num_warm_loops = %d\n", values["num_warm_loops"].int_value);
-    printf("  num_loops = %d\n\n", values["num_loops"].int_value);
+    bool preamble_printed = false;
+    size_t num_warm_loops = values["num_warm_loops"].int_value;
+    size_t num_loops = values["num_loops"].int_value;
     for (std::vector<BandwidthBenchmark*>::iterator it = bench_objs.begin();
          it != bench_objs.end(); ++it) {
-        (*it)->set_num_warm_loops(values["num_warm_loops"].int_value);
-        (*it)->set_num_loops(values["num_loops"].int_value);
+        if (!(*it)->canRun()) {
+            continue;
+        }
+        if (!(*it)->setSize(values["num_warm_loops"].int_value)) {
+            printf("Failed creating buffer for bandwidth test.\n");
+            return false;
+        }
+        if (num_warm_loops) {
+            (*it)->set_num_warm_loops(num_warm_loops);
+        }
+        if (num_loops) {
+            (*it)->set_num_loops(num_loops);
+        }
+        if (!preamble_printed) {
+            preamble_printed = true;
+            printf("Benchmarking %s bandwidth\n", name);
+            printf("  size = %d\n", (*it)->size());
+            printf("  num_warm_loops = %d\n", (*it)->num_warm_loops());
+            printf("  num_loops = %d\n\n", (*it)->num_loops());
+        }
         (*it)->run();
-        printf("  Copy bandwidth with %s: %0.2f MB/s\n", (*it)->getName(),
+        printf("  %s bandwidth with %s: %0.2f MB/s\n", name, (*it)->getName(),
                (*it)->mb_per_sec());
     }
 
-    return 0;
+    return true;
 }
 
-int write_bandwidth(int argc, char** argv) {
-    arg_t values;
-    values["size"].int_value = 0;
-    values["num_loops"].int_value = BandwidthBenchmark::DEFAULT_NUM_LOOPS;
-    values["num_warm_loops"].int_value = BandwidthBenchmark::DEFAULT_NUM_WARM_LOOPS;
-    if (!processBandwidthOptions(argc, argv, bandwidth_opts, &values)) {
+int copy_bandwidth(int argc, char** argv) {
+    std::vector<BandwidthBenchmark*> bench_objs;
+    bench_objs.push_back(new CopyLdrdStrdBenchmark());
+    bench_objs.push_back(new CopyLdmiaStmiaBenchmark());
+    bench_objs.push_back(new CopyVldVstBenchmark());
+    bench_objs.push_back(new CopyVldmiaVstmiaBenchmark());
+    bench_objs.push_back(new MemcpyBenchmark());
+
+    if (!run_bandwidth_benchmark(argc, argv, "copy", bench_objs)) {
         return -1;
     }
+    return 0;
+}
 
-    size_t size = values["size"].int_value;
-    if ((size % 64) != 0) {
-        printf("The size value must be a multiple of 64.\n");
-        return 1;
-    }
+int write_bandwidth(int argc, char** argv) {
+    std::vector<BandwidthBenchmark*> bench_objs;
+    bench_objs.push_back(new WriteStrdBenchmark());
+    bench_objs.push_back(new WriteStmiaBenchmark());
+    bench_objs.push_back(new WriteVstBenchmark());
+    bench_objs.push_back(new WriteVstmiaBenchmark());
+    bench_objs.push_back(new MemsetBenchmark());
 
-    if (setpriority(PRIO_PROCESS, 0, -20)) {
-        perror("Unable to raise priority of process.");
+    if (!run_bandwidth_benchmark(argc, argv, "write", bench_objs)) {
         return -1;
     }
 
+    return 0;
+}
+
+int read_bandwidth(int argc, char** argv) {
     std::vector<BandwidthBenchmark*> bench_objs;
-    bench_objs.push_back(new WriteStrdBenchmark(size));
-    bench_objs.push_back(new WriteStmiaBenchmark(size));
-    bench_objs.push_back(new WriteVstBenchmark(size));
-    bench_objs.push_back(new WriteVstmiaBenchmark(size));
-    bench_objs.push_back(new MemsetBenchmark(size));
-
-    printf("Benchmarking write bandwidth\n");
-    printf("  size = %d\n", bench_objs[0]->size());
-    printf("  num_warm_loops = %d\n", values["num_warm_loops"].int_value);
-    printf("  num_loops = %d\n\n", values["num_loops"].int_value);
-    for (std::vector<BandwidthBenchmark*>::iterator it = bench_objs.begin();
-         it != bench_objs.end(); ++it) {
-        (*it)->set_num_warm_loops(values["num_warm_loops"].int_value);
-        (*it)->set_num_loops(values["num_loops"].int_value);
-        (*it)->run();
-        printf("  Write bandwidth with %s: %0.2f MB/s\n", (*it)->getName(),
-               (*it)->mb_per_sec());
-    }
+    bench_objs.push_back(new ReadLdrdBenchmark());
+    bench_objs.push_back(new ReadLdmiaBenchmark());
+    bench_objs.push_back(new ReadVldBenchmark());
+    bench_objs.push_back(new ReadVldmiaBenchmark());
 
+    if (!run_bandwidth_benchmark(argc, argv, "read", bench_objs)) {
+        return -1;
+    }
     return 0;
 }
diff --git a/tests/memtest/bandwidth.h b/tests/memtest/bandwidth.h
index cef5cdd5..414977b9 100644
--- a/tests/memtest/bandwidth.h
+++ b/tests/memtest/bandwidth.h
@@ -22,13 +22,20 @@
 // Bandwidth Class definitions.
 class BandwidthBenchmark {
 public:
-    BandwidthBenchmark(size_t size)
-        : _size(size),
+    BandwidthBenchmark()
+        : _size(0),
           _num_warm_loops(DEFAULT_NUM_WARM_LOOPS),
           _num_loops(DEFAULT_NUM_LOOPS) {}
     virtual ~BandwidthBenchmark() {}
 
-    void run() {
+    bool run() {
+        if (_size == 0) {
+            return false;
+        }
+        if (!canRun()) {
+            return false;
+        }
+
         bench(_num_warm_loops);
 
         nsecs_t t = system_time();
@@ -36,12 +43,28 @@ public:
         t = system_time() - t;
 
         _mb_per_sec = (_size*(_num_loops/_BYTES_PER_MB))/(t/_NUM_NS_PER_SEC);
+
+        return true;
     }
 
+    bool canRun() { return !usesNeon() || isNeonSupported(); }
+
+    virtual bool setSize(size_t size) = 0;
+
     virtual const char *getName() = 0;
 
     virtual bool verify() = 0;
 
+    virtual bool usesNeon() { return false; }
+
+    bool isNeonSupported() {
+#if defined(__ARM_NEON__)
+        return true;
+#else
+        return false;
+#endif
+    }
+
     // Accessors/mutators.
     double mb_per_sec() { return _mb_per_sec; }
     size_t num_warm_loops() { return _num_warm_loops; }
@@ -73,22 +96,45 @@ private:
 
 class CopyBandwidthBenchmark : public BandwidthBenchmark {
 public:
-    CopyBandwidthBenchmark(size_t size) : BandwidthBenchmark(size) {
-        if (_size == 0) {
+    CopyBandwidthBenchmark() : BandwidthBenchmark(), _src(NULL), _dst(NULL) { }
+
+    bool setSize(size_t size) {
+        if (_src) {
+           free(_src);
+        }
+        if (_dst) {
+            free(_dst);
+        }
+
+        if (size == 0) {
             _size = DEFAULT_COPY_SIZE;
+        } else {
+            _size = size;
         }
+
         _src = reinterpret_cast<char*>(memalign(64, _size));
         if (!_src) {
-          perror("Failed to allocate memory for test.");
-          exit(1);
+            perror("Failed to allocate memory for test.");
+            return false;
         }
         _dst = reinterpret_cast<char*>(memalign(64, _size));
         if (!_dst) {
-          perror("Failed to allocate memory for test.");
-          exit(1);
+            perror("Failed to allocate memory for test.");
+            return false;
+        }
+
+        return true;
+    }
+    virtual ~CopyBandwidthBenchmark() {
+        if (_src) {
+            free(_src);
+            _src = NULL;
+        }
+        if (_dst) {
+            free(_dst);
+            _dst = NULL;
         }
     }
-    virtual ~CopyBandwidthBenchmark() { free(_src); free(_dst); }
 
     bool verify() {
         memset(_src, 0x23, _size);
@@ -120,7 +166,7 @@ protected:
 
 class CopyLdrdStrdBenchmark : public CopyBandwidthBenchmark {
 public:
-    CopyLdrdStrdBenchmark(size_t size) : CopyBandwidthBenchmark(size) { }
+    CopyLdrdStrdBenchmark() : CopyBandwidthBenchmark() { }
     virtual ~CopyLdrdStrdBenchmark() {}
 
     const char *getName() { return "ldrd/strd"; }
@@ -174,7 +220,7 @@ protected:
 
 class CopyLdmiaStmiaBenchmark : public CopyBandwidthBenchmark {
 public:
-    CopyLdmiaStmiaBenchmark(size_t size) : CopyBandwidthBenchmark(size) { }
+    CopyLdmiaStmiaBenchmark() : CopyBandwidthBenchmark() { }
     virtual ~CopyLdmiaStmiaBenchmark() {}
 
     const char *getName() { return "ldmia/stmia"; }
@@ -213,7 +259,7 @@ protected:
 
 class CopyVldVstBenchmark : public CopyBandwidthBenchmark {
 public:
-    CopyVldVstBenchmark(size_t size) : CopyBandwidthBenchmark(size) { }
+    CopyVldVstBenchmark() : CopyBandwidthBenchmark() { }
     virtual ~CopyVldVstBenchmark() {}
 
     const char *getName() { return "vld/vst"; }
@@ -252,7 +298,7 @@ protected:
 
 class CopyVldmiaVstmiaBenchmark : public CopyBandwidthBenchmark {
 public:
-    CopyVldmiaVstmiaBenchmark(size_t size) : CopyBandwidthBenchmark(size) { }
+    CopyVldmiaVstmiaBenchmark() : CopyBandwidthBenchmark() { }
     virtual ~CopyVldmiaVstmiaBenchmark() {}
 
     const char *getName() { return "vldmia/vstmia"; }
@@ -289,7 +335,7 @@ protected:
 
 class MemcpyBenchmark : public CopyBandwidthBenchmark {
 public:
-    MemcpyBenchmark(size_t size) : CopyBandwidthBenchmark(size) { }
+    MemcpyBenchmark() : CopyBandwidthBenchmark() { }
     virtual ~MemcpyBenchmark() {}
 
     const char *getName() { return "memcpy"; }
@@ -302,21 +348,50 @@ protected:
     }
 };
 
-class WriteBandwidthBenchmark : public BandwidthBenchmark {
+class SingleBufferBandwidthBenchmark : public BandwidthBenchmark {
 public:
-    WriteBandwidthBenchmark(size_t size) : BandwidthBenchmark(size) {
+    SingleBufferBandwidthBenchmark() : BandwidthBenchmark(), _buffer(NULL) { }
+    virtual ~SingleBufferBandwidthBenchmark() {
+        if (_buffer) {
+            free(_buffer);
+            _buffer = NULL;
+        }
+    }
+
+    bool setSize(size_t size) {
+        if (_buffer) {
+            free(_buffer);
+            _buffer = NULL;
+        }
+
         if (_size == 0) {
-            _size = DEFAULT_WRITE_SIZE;
+            _size = DEFAULT_SINGLE_BUFFER_SIZE;
+        } else {
+            _size = size;
         }
 
         _buffer = reinterpret_cast<char*>(memalign(64, _size));
         if (!_buffer) {
-          perror("Failed to allocate memory for test.");
-          exit(1);
+            perror("Failed to allocate memory for test.");
+            return false;
         }
         memset(_buffer, 0, _size);
+
+        return true;
     }
-    virtual ~WriteBandwidthBenchmark() { free(_buffer); }
+
+    bool verify() { return true; }
+
+protected:
+    char *_buffer;
+
+    static const unsigned int DEFAULT_SINGLE_BUFFER_SIZE = 16000;
+};
+
+class WriteBandwidthBenchmark : public SingleBufferBandwidthBenchmark {
+public:
+    WriteBandwidthBenchmark() : SingleBufferBandwidthBenchmark() { }
+    virtual ~WriteBandwidthBenchmark() { }
 
     bool verify() {
         memset(_buffer, 0, _size);
@@ -339,16 +414,11 @@ public:
 
         return true;
     }
-
-protected:
-    char *_buffer;
-
-    static const unsigned int DEFAULT_WRITE_SIZE = 16000;
 };
 
 class WriteStrdBenchmark : public WriteBandwidthBenchmark {
 public:
-    WriteStrdBenchmark(size_t size) : WriteBandwidthBenchmark(size) { }
+    WriteStrdBenchmark() : WriteBandwidthBenchmark() { }
     virtual ~WriteStrdBenchmark() {}
 
     const char *getName() { return "strd"; }
@@ -392,7 +462,7 @@ protected:
 
 class WriteStmiaBenchmark : public WriteBandwidthBenchmark {
 public:
-    WriteStmiaBenchmark(size_t size) : WriteBandwidthBenchmark(size) { }
+    WriteStmiaBenchmark() : WriteBandwidthBenchmark() { }
     virtual ~WriteStmiaBenchmark() {}
 
     const char *getName() { return "stmia"; }
@@ -437,11 +507,13 @@ protected:
 
 class WriteVstBenchmark : public WriteBandwidthBenchmark {
 public:
-    WriteVstBenchmark(size_t size) : WriteBandwidthBenchmark(size) { }
+    WriteVstBenchmark() : WriteBandwidthBenchmark() { }
     virtual ~WriteVstBenchmark() {}
 
     const char *getName() { return "vst"; }
 
+    bool usesNeon() { return true; }
+
 protected:
     // Write a given value using vst.
     void bench(size_t num_loops) {
@@ -480,11 +552,13 @@ protected:
 
 class WriteVstmiaBenchmark : public WriteBandwidthBenchmark {
 public:
-    WriteVstmiaBenchmark(size_t size) : WriteBandwidthBenchmark(size) { }
+    WriteVstmiaBenchmark() : WriteBandwidthBenchmark() { }
     virtual ~WriteVstmiaBenchmark() {}
 
     const char *getName() { return "vstmia"; }
 
+    bool usesNeon() { return true; }
+
 protected:
     // Write a given value using vstmia.
     void bench(size_t num_loops) {
@@ -523,7 +597,7 @@ protected:
 
 class MemsetBenchmark : public WriteBandwidthBenchmark {
 public:
-    MemsetBenchmark(size_t size) : WriteBandwidthBenchmark(size) { }
+    MemsetBenchmark() : WriteBandwidthBenchmark() { }
     virtual ~MemsetBenchmark() {}
 
     const char *getName() { return "memset"; }
@@ -536,4 +610,152 @@ protected:
     }
 };
 
+class ReadLdrdBenchmark : public SingleBufferBandwidthBenchmark {
+public:
+    ReadLdrdBenchmark() : SingleBufferBandwidthBenchmark() { }
+    virtual ~ReadLdrdBenchmark() {}
+
+    const char *getName() { return "ldrd"; }
+
+protected:
+    // Write a given value using strd.
+    void bench(size_t num_loops) {
+        asm volatile(
+            "stmfd sp!, {r0,r1,r2,r3,r4,r5}\n"
+
+            "mov r0, %0\n"
+            "mov r1, %1\n"
+            "mov r2, %2\n"
+
+            "0:\n"
+            "mov r3, r1, lsr #5\n"
+
+            "1:\n"
+            "subs r3, r3, #1\n"
+            "ldrd r4, r5, [r0]\n"
+            "ldrd r4, r5, [r0, #8]\n"
+            "ldrd r4, r5, [r0, #16]\n"
+            "ldrd r4, r5, [r0, #24]\n"
+            "add  r0, r0, #32\n"
+            "bgt 1b\n"
+
+            "sub r0, r0, r1\n"
+            "subs r2, r2, #1\n"
+            "bgt 0b\n"
+
+            "ldmfd sp!, {r0,r1,r2,r3,r4,r5}\n"
+          :: "r" (_buffer), "r" (_size), "r" (num_loops) : "r0", "r1", "r2");
+    }
+};
+
+class ReadLdmiaBenchmark : public SingleBufferBandwidthBenchmark {
+public:
+    ReadLdmiaBenchmark() : SingleBufferBandwidthBenchmark() { }
+    virtual ~ReadLdmiaBenchmark() {}
+
+    const char *getName() { return "ldmia"; }
+
+protected:
+      // Write a given value using stmia.
+      void bench(size_t num_loops) {
+          asm volatile(
+              "stmfd sp!, {r0,r1,r2,r3,r4,r5,r6,r7,r8,r9,r10,r11}\n"
+
+              "mov r0, %0\n"
+              "mov r1, %1\n"
+              "mov r2, %2\n"
+
+              "0:\n"
+              "mov r3, r1, lsr #5\n"
+
+              "1:\n"
+              "subs r3, r3, #1\n"
+              "ldmia r0!, {r4, r5, r6, r7, r8, r9, r10, r11}\n"
+              "bgt 1b\n"
+
+              "sub r0, r0, r1\n"
+              "subs r2, r2, #1\n"
+              "bgt 0b\n"
+
+              "ldmfd sp!, {r0,r1,r2,r3,r4,r5,r6,r7,r8,r9,r10,r11}\n"
+        :: "r" (_buffer), "r" (_size), "r" (num_loops) : "r0", "r1", "r2");
+    }
+};
+
+class ReadVldBenchmark : public SingleBufferBandwidthBenchmark {
+public:
+    ReadVldBenchmark() : SingleBufferBandwidthBenchmark() { }
+    virtual ~ReadVldBenchmark() {}
+
+    const char *getName() { return "vld"; }
+
+    bool usesNeon() { return true; }
+
+protected:
+    // Write a given value using vst.
+    void bench(size_t num_loops) {
+#if defined(__ARM_NEON__)
+        asm volatile(
+            "stmfd sp!, {r0,r1,r2,r3}\n"
+
+            "mov r0, %0\n"
+            "mov r1, %1\n"
+            "mov r2, %2\n"
+
+            "0:\n"
+            "mov r3, r1, lsr #5\n"
+
+            "1:\n"
+            "subs r3, r3, #1\n"
+            "vld1.8 {d0-d3}, [r0:128]!\n"
+            "bgt 1b\n"
+
+            "sub r0, r0, r1\n"
+            "subs r2, r2, #1\n"
+            "bgt 0b\n"
+
+            "ldmfd sp!, {r0,r1,r2,r3}\n"
+        :: "r" (_buffer), "r" (_size), "r" (num_loops) : "r0", "r1", "r2");
+#endif
+    }
+};
+
+class ReadVldmiaBenchmark : public SingleBufferBandwidthBenchmark {
+public:
+    ReadVldmiaBenchmark() : SingleBufferBandwidthBenchmark() { }
+    virtual ~ReadVldmiaBenchmark() {}
+
+    const char *getName() { return "vldmia"; }
+
+    bool usesNeon() { return true; }
+
+protected:
+    // Write a given value using vstmia.
+    void bench(size_t num_loops) {
+#if defined(__ARM_NEON__)
+        asm volatile(
+            "stmfd sp!, {r0,r1,r2,r3}\n"
+
+            "mov r0, %0\n"
+            "mov r1, %1\n"
+            "mov r2, %2\n"
+
+            "0:\n"
+            "mov r3, r1, lsr #5\n"
+
+            "1:\n"
+            "subs r3, r3, #1\n"
+            "vldmia r0!, {d0-d3}\n"
+            "bgt 1b\n"
+
+            "sub r0, r0, r1\n"
+            "subs r2, r2, #1\n"
+            "bgt 0b\n"
+
+            "ldmfd sp!, {r0,r1,r2,r3}\n"
+        :: "r" (_buffer), "r" (_size), "r" (num_loops) : "r0", "r1", "r2");
+#endif
+    }
+};
+
 #endif  // __BANDWIDTH_H__
diff --git a/tests/memtest/memtest.cpp b/tests/memtest/memtest.cpp
index 0d76b45d..86ec9a41 100644
--- a/tests/memtest/memtest.cpp
+++ b/tests/memtest/memtest.cpp
@@ -41,14 +41,17 @@ static void usage(char* p) {
            "<test> is one of the following:\n"
            "  copy_bandwidth [--size BYTES_TO_COPY]\n"
            "  write_bandwidth [--size BYTES_TO_WRITE]\n"
+           "  read_bandwidth [--size BYTES_TO_COPY]\n"
            "  per_core_bandwidth [--size BYTES]\n"
            "    --type copy_ldrd_strd | copy_ldmia_stmia | copy_vld_vst |\n"
            "           copy_vldmia_vstmia | memcpy | write_strd | write_stmia |\n"
-           "           write_vst | write_vstmia | memset\n"
+           "           write_vst | write_vstmia | memset | read_ldrd |\n"
+           "           read_ldmia | read_vld | read_vldmia\n"
            "  multithread_bandwidth [--size BYTES]\n"
            "    --type copy_ldrd_strd | copy_ldmia_stmia | copy_vld_vst |\n"
            "           copy_vldmia_vstmia | memcpy | write_strd | write_stmia |\n"
-           "           write_vst | write_vstmia | memset\n"
+           "           write_vst | write_vstmia | memset | read_ldrd |\n"
+           "           read_ldmia | read_vld | read_vldmia\n"
            "    --num_threads NUM_THREADS_TO_RUN\n"
            "  malloc [fill]\n"
            "  madvise\n"
@@ -61,6 +64,7 @@ static void usage(char* p) {
 
 int copy_bandwidth(int argc, char** argv);
 int write_bandwidth(int argc, char** argv);
+int read_bandwidth(int argc, char** argv);
 int per_core_bandwidth(int argc, char** argv);
 int multithread_bandwidth(int argc, char** argv);
 int malloc_test(int argc, char** argv);
@@ -83,7 +87,8 @@ function_t function_table[] = {
     { "crawl", crawl_test },
     { "fp", fp_test },
     { "copy_bandwidth", copy_bandwidth },
-    { "write_bandwidth", write_bandwidth},
+    { "write_bandwidth", write_bandwidth },
+    { "read_bandwidth", read_bandwidth },
     { "per_core_bandwidth", per_core_bandwidth },
     { "multithread_bandwidth", multithread_bandwidth },
 };