FPGA gzip area fix (oneapi-src#639)

* Fix segfaults in GZIP design without adding predication to LSU in
kernel, which caused an area increase.

* GZIP seed update

* Seed update for low latency version

Author: tyoungsc

DirectProgramming/DPC++FPGA/ReferenceDesigns/gzip/src/CMakeLists.txt

@@ -35,13 +35,13 @@ endif()
 if(FPGA_BOARD MATCHES ".*a10.*")
     # A10 parameters
     set(NUM_ENGINES 1)
-    set(LL_SEED "-Xsseed=1")
-    set(HIGH_BW_SEED "-Xsseed=6")
+    set(LL_SEED "-Xsseed=4")
+    set(HIGH_BW_SEED "-Xsseed=4")
     set(NUM_REORDER "")
 elseif(FPGA_BOARD MATCHES ".*s10.*")
     # S10 parameters
     set(NUM_ENGINES 2)
-    set(LL_SEED "-Xsseed=2")
+    set(LL_SEED "-Xsseed=6")
     set(HIGH_BW_SEED "-Xsseed=2")
     set(NUM_REORDER "-Xsnum-reorder=6")
 elseif(FPGA_BOARD MATCHES ".*agilex.*")

DirectProgramming/DPC++FPGA/ReferenceDesigns/gzip/src/gzip.cpp

@@ -277,6 +277,10 @@ int CompressFile(queue &q, std::string &input_file, std::vector<std::string> out
 
   int buffers_count = iterations;
 
+  // padding for the input and output buffers to deal with granularity of
+  // kernel reads and writes
+  constexpr size_t kInOutPadding = 16 * kVec;
+
   // Create an array of kernel info structures and create buffers for kernel
   // input/output. The buffers are re-used between iterations, but enough 
   // disjoint buffers are created to support double-buffering.
@@ -292,9 +296,10 @@ int CompressFile(queue &q, std::string &input_file, std::vector<std::string> out
       kinfo[eng][i].file_size = isz;
       // Allocating slightly larger buffers (+ 16 * kVec) to account for
       // granularity of kernel writes
-      int outputSize = kinfo[eng][i].file_size + 16 * kVec < kMinBufferSize
-                           ? kMinBufferSize
-                           : kinfo[eng][i].file_size + 16 * kVec;
+      int outputSize =
+          ((isz + kInOutPadding) < kMinBufferSize) ? kMinBufferSize
+                                                   : (isz + kInOutPadding);
+      const size_t input_alloc_size = isz + kInOutPadding;
 
       // Pre-pin buffer using malloc_host() to improve DMA bandwidth.
       if (i >= 3) {
@@ -327,7 +332,7 @@ int CompressFile(queue &q, std::string &input_file, std::vector<std::string> out
                                       : new buffer<unsigned, 1>(kMinBufferSize);
       kinfo[eng][i].pibuf = i >= 3
                                 ? kinfo[eng][i - 3].pibuf
-                                : new buffer<char, 1>(kinfo[eng][i].file_size);
+                                : new buffer<char, 1>(input_alloc_size);
       kinfo[eng][i].pobuf =
           i >= 3 ? kinfo[eng][i - 3].pobuf : new buffer<char, 1>(outputSize);
       kinfo[eng][i].pobuf_decompress = (char *)malloc(kinfo[eng][i].file_size);

DirectProgramming/DPC++FPGA/ReferenceDesigns/gzip/src/gzip_ll.cpp

@@ -306,14 +306,20 @@ int CompressFile(queue &q, std::string &input_file,
     }
   }
 
+  // padding for the input and output buffers to deal with granularity of
+  // kernel reads and writes
+  constexpr size_t kInOutPadding = 16 * kVec;
+
   // This loop allocates host-side USM buffers, to be accessed by the kernel.
   for (size_t eng = 0; eng < kNumEngines; eng++) {
     for (int i = 0; i < buffers_count; i++) {
       kinfo[eng][i].input_size = isz;
       // Allocating slightly larger buffers (+ 16 * kVec) to account for
       // granularity of kernel writes
       kinfo[eng][i].output_size =
-          isz + 16 * kVec < kMinBufferSize ? kMinBufferSize : isz + 16 * kVec;
+          ((isz + kInOutPadding) < kMinBufferSize) ? kMinBufferSize
+                                                   : (isz + kInOutPadding);
+      const size_t input_alloc_size = isz + kInOutPadding;
 
       kinfo[eng][i].last_block = true;
       kinfo[eng][i].pref_buffer = pinbuf;
@@ -351,7 +357,7 @@ int CompressFile(queue &q, std::string &input_file,
                           // since the buffers get subsequently reused.
         for (int b = 0; b < BATCH_SIZE; b++) {
           kinfo[eng][i].pibuf_ptr_array[b] =
-              alloc_char.allocate(kinfo[eng][i].input_size * sizeof(char));
+              alloc_char.allocate(input_alloc_size * sizeof(char));
           kinfo[eng][i].pobuf_ptr_array[b] =
               alloc_char.allocate(kinfo[eng][i].output_size * sizeof(char));
           memset(kinfo[eng][i].pobuf_ptr_array[b], 0,

DirectProgramming/DPC++FPGA/ReferenceDesigns/gzip/src/gzipkernel.cpp

@@ -2063,14 +2063,11 @@ void SubmitGzipTasksSingleEngine(
 
       // load in new data
       struct LzInput in;
+      Unroller<0, kVec>::step([&](int i) { in.data[i] = acc_pibuf[inpos++]; });
       Unroller<0, kVec>::step([&](int i) {
-        // prevent out-of-bounds reads
-        in.data[i] = (inpos < accessor_isz) ? acc_pibuf[inpos++] : 0;
+        current_window[i + kVec] = in.data[i];
       });
 
-      Unroller<0, kVec>::step(
-          [&](int i) { current_window[i + kVec] = in.data[i]; });
-
       do {
         //-----------------------------
         // Prepare current window

DirectProgramming/DPC++FPGA/ReferenceDesigns/gzip/src/gzipkernel_ll.cpp

@@ -2193,8 +2193,7 @@ event SubmitLZReduction(queue &q, size_t block_size, bool last_block,
 
           // load in new data
           Unroller<0, kVec>::step([&](int i) {
-            // guarding against out-of-bounds accesses
-            in.data[i] = (inpos < accessor_isz) ? acc_pibuf[inpos++] : 0;
+            in.data[i] = acc_pibuf[inpos++];
             input_data.arr[16 * (int)crc_ch_load_upper + i] = in.data[i];
           });