Add more comments and error checking.

src/runtime/hexagon_remote/bin/src/halide_hexagon_remote.h

@@ -37,7 +37,7 @@ __QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_initialize_kernels_
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_get_symbol_v4)(halide_hexagon_remote_handle_t module_ptr, const char* name, int nameLen, halide_hexagon_remote_handle_t* sym_ptr) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_power_hvx_on)(void) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_power_hvx_off)(void) __QAIC_HEADER_ATTRIBUTE;
-__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_run_v2)(halide_hexagon_remote_handle_t module_ptr, halide_hexagon_remote_handle_t symbol, const halide_hexagon_remote_buffer* input_buffers, int input_buffersLen, halide_hexagon_remote_buffer* output_buffers, int output_buffersLen, int scalar_count, const unsigned char* small, int smallLen) __QAIC_HEADER_ATTRIBUTE;
+__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_run_v2)(halide_hexagon_remote_handle_t module_ptr, halide_hexagon_remote_handle_t symbol, const halide_hexagon_remote_buffer* input_buffers, int input_buffersLen, halide_hexagon_remote_buffer* output_buffers, int output_buffersLen, int scalar_count, const unsigned char* small_input_args, int small_input_argsLen) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_release_kernels_v2)(halide_hexagon_remote_handle_t module_ptr) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_poll_log)(char* log, int logLen, int* read_size) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_poll_profiler_state)(int* func, int* threads) __QAIC_HEADER_ATTRIBUTE;

src/runtime/hexagon_remote/bin/src/halide_hexagon_remote_skel.c

@@ -425,8 +425,8 @@ static const Parameter parameters[8] = {{SLIM_IFPTR32(0x8,0x10),{{(const uintptr
 static const Parameter* const parameterArrays[21] = {(&(parameters[5])),(&(parameters[2])),(&(parameters[2])),(&(parameters[5])),(&(parameters[2])),(&(parameters[2])),(&(parameters[5])),(&(parameters[5])),(&(parameters[2])),(&(parameters[2])),(&(parameters[3])),(&(parameters[4])),(&(parameters[5])),(&(parameters[0])),(&(parameters[2])),(&(parameters[0])),(&(parameters[1])),(&(parameters[7])),(&(parameters[7])),(&(parameters[6])),(&(parameters[7]))};
 static const Method methods[9] = {{REMOTE_SCALARS_MAKEX(0,0,0x2,0x1,0x0,0x0),0x4,0x4,3,2,(&(parameterArrays[15])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x2,0x1,0x0,0x0),0x8,0x4,4,3,(&(parameterArrays[14])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x0,0x0,0x0,0x0),0x0,0x0,0,0,0,0x0,0x0},{REMOTE_SCALARS_MAKEX(0,0,255,255,15,15),0x18,0x0,10,6,(&(parameterArrays[8])),0x4,0x1},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x4,0x0,1,1,(&(parameterArrays[1])),0x4,0x0},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x2,0x0,0x0),0x4,0x4,4,2,(&(parameterArrays[19])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x0,0x1,0x0,0x0),0x0,0x8,2,2,(&(parameterArrays[17])),0x1,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x4,0x0,1,1,(&(parameterArrays[0])),0x4,0x0},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x20,0x0,8,8,(&(parameterArrays[0])),0x4,0x0}};
 static const Method* const methodArrays[10] = {&(methods[0]),&(methods[1]),&(methods[2]),&(methods[2]),&(methods[3]),&(methods[4]),&(methods[5]),&(methods[6]),&(methods[7]),&(methods[8])};
-static const char strings[357] = "initialize_kernels_v3\0busbwUsagePercentage\0set_performance_mode\0poll_profiler_state\0release_kernels_v2\0bwMegabytesPerSec\0set_performance\0output_buffers\0mipsPerThread\0input_buffers\0power_hvx_off\0get_symbol_v4\0scalar_count\0power_hvx_on\0set_latency\0set_bus_bw\0module_ptr\0mipsTotal\0read_size\0set_mips\0poll_log\0threads\0sym_ptr\0symbol\0run_v2\0small\0func\0name\0code\0";
-static const uint16_t methodStrings[35] = {121,288,152,268,246,103,22,234,238,329,257,322,166,137,208,336,194,257,347,314,64,342,306,297,302,278,0,352,257,43,59,84,257,180,221};
+static const char strings[368] = "initialize_kernels_v3\0busbwUsagePercentage\0set_performance_mode\0poll_profiler_state\0release_kernels_v2\0bwMegabytesPerSec\0small_input_args\0set_performance\0output_buffers\0mipsPerThread\0input_buffers\0power_hvx_off\0get_symbol_v4\0scalar_count\0power_hvx_on\0set_latency\0set_bus_bw\0module_ptr\0mipsTotal\0read_size\0set_mips\0poll_log\0threads\0sym_ptr\0symbol\0run_v2\0func\0name\0code\0";
+static const uint16_t methodStrings[35] = {138,305,169,285,263,103,22,251,255,346,274,339,183,154,225,121,211,274,358,331,64,353,323,314,319,295,0,363,274,43,59,84,274,197,238};
 static const uint16_t methodStringsArrays[10] = {26,16,34,33,9,31,23,20,29,0};
 __QAIC_SLIM_EXPORT const Interface __QAIC_SLIM(halide_hexagon_remote_slim) = {10,&(methodArrays[0]),0,0,&(methodStringsArrays [0]),methodStrings,strings};
 #endif //_HALIDE_HEXAGON_REMOTE_SLIM_H

src/runtime/hexagon_remote/bin/src/halide_hexagon_remote_stub.c

@@ -425,8 +425,8 @@ static const Parameter parameters[8] = {{SLIM_IFPTR32(0x8,0x10),{{(const uintptr
 static const Parameter* const parameterArrays[21] = {(&(parameters[5])),(&(parameters[2])),(&(parameters[2])),(&(parameters[5])),(&(parameters[2])),(&(parameters[2])),(&(parameters[5])),(&(parameters[5])),(&(parameters[2])),(&(parameters[2])),(&(parameters[3])),(&(parameters[4])),(&(parameters[5])),(&(parameters[0])),(&(parameters[2])),(&(parameters[0])),(&(parameters[1])),(&(parameters[7])),(&(parameters[7])),(&(parameters[6])),(&(parameters[7]))};
 static const Method methods[9] = {{REMOTE_SCALARS_MAKEX(0,0,0x2,0x1,0x0,0x0),0x4,0x4,3,2,(&(parameterArrays[15])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x2,0x1,0x0,0x0),0x8,0x4,4,3,(&(parameterArrays[14])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x0,0x0,0x0,0x0),0x0,0x0,0,0,0,0x0,0x0},{REMOTE_SCALARS_MAKEX(0,0,255,255,15,15),0x18,0x0,10,6,(&(parameterArrays[8])),0x4,0x1},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x4,0x0,1,1,(&(parameterArrays[1])),0x4,0x0},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x2,0x0,0x0),0x4,0x4,4,2,(&(parameterArrays[19])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x0,0x1,0x0,0x0),0x0,0x8,2,2,(&(parameterArrays[17])),0x1,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x4,0x0,1,1,(&(parameterArrays[0])),0x4,0x0},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x20,0x0,8,8,(&(parameterArrays[0])),0x4,0x0}};
 static const Method* const methodArrays[10] = {&(methods[0]),&(methods[1]),&(methods[2]),&(methods[2]),&(methods[3]),&(methods[4]),&(methods[5]),&(methods[6]),&(methods[7]),&(methods[8])};
-static const char strings[357] = "initialize_kernels_v3\0busbwUsagePercentage\0set_performance_mode\0poll_profiler_state\0release_kernels_v2\0bwMegabytesPerSec\0set_performance\0output_buffers\0mipsPerThread\0input_buffers\0power_hvx_off\0get_symbol_v4\0scalar_count\0power_hvx_on\0set_latency\0set_bus_bw\0module_ptr\0mipsTotal\0read_size\0set_mips\0poll_log\0threads\0sym_ptr\0symbol\0run_v2\0small\0func\0name\0code\0";
-static const uint16_t methodStrings[35] = {121,288,152,268,246,103,22,234,238,329,257,322,166,137,208,336,194,257,347,314,64,342,306,297,302,278,0,352,257,43,59,84,257,180,221};
+static const char strings[368] = "initialize_kernels_v3\0busbwUsagePercentage\0set_performance_mode\0poll_profiler_state\0release_kernels_v2\0bwMegabytesPerSec\0small_input_args\0set_performance\0output_buffers\0mipsPerThread\0input_buffers\0power_hvx_off\0get_symbol_v4\0scalar_count\0power_hvx_on\0set_latency\0set_bus_bw\0module_ptr\0mipsTotal\0read_size\0set_mips\0poll_log\0threads\0sym_ptr\0symbol\0run_v2\0func\0name\0code\0";
+static const uint16_t methodStrings[35] = {138,305,169,285,263,103,22,251,255,346,274,339,183,154,225,121,211,274,358,331,64,353,323,314,319,295,0,363,274,43,59,84,274,197,238};
 static const uint16_t methodStringsArrays[10] = {26,16,34,33,9,31,23,20,29,0};
 __QAIC_SLIM_EXPORT const Interface __QAIC_SLIM(halide_hexagon_remote_slim) = {10,&(methodArrays[0]),0,0,&(methodStringsArrays [0]),methodStrings,strings};
 #endif //_HALIDE_HEXAGON_REMOTE_SLIM_H
@@ -714,9 +714,9 @@ static __inline int _stub_method_3(remote_handle _handle, uint32_t _mid, uint32_
    _allocator_deinit(_al);
    return _nErr;
 }
-__QAIC_STUB_EXPORT int __QAIC_STUB(halide_hexagon_remote_run_v2)(halide_hexagon_remote_handle_t module_ptr, halide_hexagon_remote_handle_t symbol, const halide_hexagon_remote_buffer* input_buffers, int input_buffersLen, halide_hexagon_remote_buffer* output_buffers, int output_buffersLen, int scalar_count, const unsigned char* small, int smallLen) __QAIC_STUB_ATTRIBUTE {
+__QAIC_STUB_EXPORT int __QAIC_STUB(halide_hexagon_remote_run_v2)(halide_hexagon_remote_handle_t module_ptr, halide_hexagon_remote_handle_t symbol, const halide_hexagon_remote_buffer* input_buffers, int input_buffersLen, halide_hexagon_remote_buffer* output_buffers, int output_buffersLen, int scalar_count, const unsigned char* small_input_args, int small_input_argsLen) __QAIC_STUB_ATTRIBUTE {
    uint32_t _mid = 4;
-   return _stub_method_3(_halide_hexagon_remote_handle(), _mid, (uint32_t*)&module_ptr, (uint32_t*)&symbol, (void**)&input_buffers, (uint32_t*)&input_buffersLen, (void**)&output_buffers, (uint32_t*)&output_buffersLen, (uint32_t*)&scalar_count, (char**)&small, (uint32_t*)&smallLen);
+   return _stub_method_3(_halide_hexagon_remote_handle(), _mid, (uint32_t*)&module_ptr, (uint32_t*)&symbol, (void**)&input_buffers, (uint32_t*)&input_buffersLen, (void**)&output_buffers, (uint32_t*)&output_buffersLen, (uint32_t*)&scalar_count, (char**)&small_input_args, (uint32_t*)&small_input_argsLen);
 }
 static __inline int _stub_method_4(remote_handle _handle, uint32_t _mid, uint32_t _in0[1]) {
    remote_arg _pra[1];

src/runtime/hexagon_remote/halide_hexagon_remote.idl

@@ -23,7 +23,7 @@ interface halide_hexagon_remote
     long run_v2(in handle_t module_ptr, in handle_t symbol,
                 in sequence<buffer> input_buffers,
                 rout sequence<buffer> output_buffers,
-                in long scalar_count, in buffer small);
+                in long scalar_count, in buffer small_input_args);
 
     // Routine to clean up a module on the remote side.
     long release_kernels_v2(in handle_t module_ptr);

src/runtime/hexagon_remote/halide_remote.cpp

@@ -316,7 +316,8 @@ int halide_hexagon_remote_get_symbol_v4(handle_t module_ptr, const char* name, i
 int halide_hexagon_remote_run_v2(handle_t module_ptr, handle_t function,
                                  const buffer *input_buffersPtrs, int input_buffersLen,
                                  buffer *output_buffersPtrs, int output_buffersLen,
-                                 int scalar_count, const unsigned char *small, int smallLen) {
+                                 int scalar_count,
+                                 const unsigned char *small_input_args, int small_input_argsLen) {
     // Get a pointer to the argv version of the pipeline.
     pipeline_argv_t pipeline = reinterpret_cast<pipeline_argv_t>(function);
 
@@ -332,6 +333,8 @@ int halide_hexagon_remote_run_v2(handle_t module_ptr, handle_t function,
     void **args = (void **)__builtin_alloca((input_buffersLen + scalar_count + output_buffersLen) * sizeof(void *));
     buffer_t *buffers = (buffer_t *)__builtin_alloca((input_buffersLen + output_buffersLen) * sizeof(buffer_t));
 
+    const unsigned char *small_input_args_end = small_input_args + small_input_argsLen;
+
     void **next_arg = &args[0];
     buffer_t *next_buffer_t = &buffers[0];
     // Input buffers come first.
@@ -340,7 +343,12 @@ int halide_hexagon_remote_run_v2(handle_t module_ptr, handle_t function,
             // This buffer is passed directly.
             next_buffer_t->host = input_buffersPtrs[i].data;
         } else {
-            next_buffer_t->host = const_cast<unsigned char *>(read_buffer(small));
+            next_buffer_t->host = const_cast<unsigned char *>(read_buffer(small_input_args));
+            if (small_input_args > small_input_args_end) {
+                log_printf("Input buffer %d read past the end of small_input_args [%p, %p)\n",
+                           i, small_input_args - small_input_args_end);
+                return halide_error_code_access_out_of_bounds;
+            }
         }
         *next_arg = next_buffer_t;
     }
@@ -352,7 +360,12 @@ int halide_hexagon_remote_run_v2(handle_t module_ptr, handle_t function,
     }
     // Input scalars are last.
     for (int i = 0; i < scalar_count; i++, next_arg++) {
-        *next_arg = const_cast<unsigned char *>(read_buffer(small));
+        *next_arg = const_cast<unsigned char *>(read_buffer(small_input_args));
+        if (small_input_args > small_input_args_end) {
+            log_printf("Input scalar %d read past the end of small_input_args by %d bytes\n",
+                       i, small_input_args - small_input_args_end);
+            return halide_error_code_access_out_of_bounds;
+        }
     }
 
     // Prior to running the pipeline, power HVX on (if it was not already on).

src/runtime/hexagon_remote/host_shim.cpp

@@ -16,27 +16,29 @@ int halide_hexagon_remote_run(handle_t module_ptr, handle_t function,
     const int buffer_alignment = 128;
     const int scalar_alignment = 4;
 
-    int small_size = 0;
+    int small_input_args_size = 0;
 
-    // Add up the number and size of the small input buffers.
+    // Add up the number and size of the small input buffers. Buffers
+    // are considered small if they are less than
+    // min_input_buffer_size bytes.
     const int min_input_buffer_size = 4096;
     for (int i = 0; i < input_buffersLen; i++) {
         if (input_buffersPtrs[i].dataLen < min_input_buffer_size) {
-            small_size += packed_buffer_size(input_buffersPtrs[i].dataLen, buffer_alignment);
+            small_input_args_size += packed_buffer_size(input_buffersPtrs[i].dataLen, buffer_alignment);
         }
     }
 
     // Add up the number and size of the scalars.
     for (int i = 0; i < input_scalarsLen; i++) {
-        small_size += packed_buffer_size(input_scalarsPtrs[i].dataLen, scalar_alignment);
+        small_input_args_size += packed_buffer_size(input_scalarsPtrs[i].dataLen, scalar_alignment);
     }
 
     // Allocate a buffer to hold the packed small buffers and scalars.
-    unsigned char *small = (unsigned char *)__builtin_alloca(small_size + buffer_alignment);
-    small = (unsigned char *)(((uintptr_t)small + buffer_alignment - 1) & ~(buffer_alignment - 1));
+    unsigned char *small_input_args = (unsigned char *)__builtin_alloca(small_input_args_size + buffer_alignment);
+    small_input_args = (unsigned char *)(((uintptr_t)small_input_args + buffer_alignment - 1) & ~(buffer_alignment - 1));
 
     // Pack up the small buffers.
-    unsigned char *write = small;
+    unsigned char *write = small_input_args;
     for (int i = 0; i < input_buffersLen; i++) {
         if (input_buffersPtrs[i].dataLen < min_input_buffer_size) {
             write_buffer(write, input_buffersPtrs[i].data, input_buffersPtrs[i].dataLen, buffer_alignment);
@@ -56,7 +58,7 @@ int halide_hexagon_remote_run(handle_t module_ptr, handle_t function,
     return halide_hexagon_remote_run_v2(module_ptr, function,
                                         input_buffersPtrs, input_buffersLen,
                                         output_buffersPtrs, output_buffersLen,
-                                        input_scalarsLen, small, small_size);
+                                        input_scalarsLen, small_input_args, small_input_args_size);
 }
 
 }  // extern "C"

src/runtime/hexagon_remote/pack_buffer.h

@@ -3,6 +3,9 @@
 
 #include <stdint.h>
 
+// This header implements a simple buffer packing protocol for packing
+// and unpacking scalars and buffers into a larger buffer.
+
 template <typename T>
 void write_scalar(unsigned char *&write, const T& value) {
     *(T *)write = value;
@@ -34,6 +37,8 @@ inline uint32_t aligned_offset(const unsigned char *start, uint32_t offset, uint
     return x - (uintptr_t)start;
 }
 
+// Buffers are stored with two offsets, indicating the start and end
+// of the buffer from the start.
 inline void write_buffer(unsigned char *&write, const unsigned char *buffer, uint32_t size, uint32_t alignment) {
     unsigned char *start = write;
     uint32_t header_size = sizeof(uint32_t) * 2;