Support for Custom Pitch for gfx103x

Signed-off-by: Mike Li <Tianxinmike.Li@amd.com>
Change-Id: Ica83dff8bb382637010396781190f585754bd150


[ROCm/ROCR-Runtime commit: 93609fd3d4]

projects/rocr-runtime/runtime/hsa-runtime/image/addrlib/src/core/addrlib2.cpp

@@ -314,8 +314,6 @@ ADDR_E_RETURNCODE Lib::ComputeSurfaceInfo(
         }
     }
 
-    ADDR_ASSERT(pOut->surfSize != 0);
-
     ValidBaseAlignments(pOut->baseAlign);
 
     return returnCode;

projects/rocr-runtime/runtime/hsa-runtime/image/image_manager_nv.cpp

@@ -484,7 +484,11 @@ hsa_status_t ImageManagerNv::PopulateImageSrd(Image& image) const {
     word4.f.DEPTH =
         (image_array) // Doesn't hurt but isn't array_size already >0?
             ? std::max(image.desc.array_size, static_cast<size_t>(1)) - 1
-            : (image_3d) ? image.desc.depth - 1 : out.pitch - 1;
+            : (image_3d) ? image.desc.depth - 1 : 0;
+    uint32_t minor_ver = MinorVerFromDevID(chip_id_);
+    // For 1d, 2d and 2d-msaa in gfx1030 and beyond this is pitch-1
+    if ((minor_ver >= 3) && !image_array && !image_3d)
+      word4.f.PITCH = out.pitch - 1;
 
     word5.val = 0;
     word6.val = 0;
@@ -630,6 +634,7 @@ uint32_t ImageManagerNv::GetAddrlibSurfaceInfoNv(
   const uint32_t num_slice = static_cast<uint32_t>(
       std::max(kMinNumSlice, std::max(desc.array_size, desc.depth)));
 
+  uint32_t minor_ver = MinorVerFromDevID(chip_id_);
   ADDR2_COMPUTE_SURFACE_INFO_INPUT in = {0};
   in.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_INPUT);
   in.format = addrlib_format;
@@ -637,6 +642,9 @@ uint32_t ImageManagerNv::GetAddrlibSurfaceInfoNv(
   in.width = width;
   in.height = height;
   in.numSlices = num_slice;
+  // Custom Pitch is supported in gfx1030 and beyond
+  if (minor_ver >= 3)
+    in.pitchInElement = image_data_row_pitch / image_prop.element_size;
   switch (desc.geometry) {
     case HSA_EXT_IMAGE_GEOMETRY_1D:
     case HSA_EXT_IMAGE_GEOMETRY_1DB:

projects/rocr-runtime/runtime/hsa-runtime/image/resource_nv.h

@@ -300,21 +300,34 @@ union SQ_IMG_RSRC_WORD3 {
 #define SQ_IMG_RSC_WRD4_REG_SZ 32
 #define SQ_IMG_RSC_WRD4_DEPTH_SZ    13
 #define SQ_IMG_RSC_WRD4_BASE_ARR_SZ 13
-struct sq_img_rsrc_word4_t {
+#define SQ_IMG_RSC_WRD4_PITCH_SZ 14
+union sq_img_rsrc_word4_t {
+  struct {
 #if defined(LITTLEENDIAN_CPU)
-  // For arrays this is last slice in view, for 3D this is depth-1, For remaining this is pitch-1
-  unsigned int DEPTH      : SQ_IMG_RSC_WRD4_DEPTH_SZ;
-  unsigned int            : 1; //Pitch[13] in gfx1030
-  unsigned int            : 2;
-  unsigned int BASE_ARRAY : SQ_IMG_RSC_WRD4_BASE_ARR_SZ;
-  unsigned int            : 3;
+    // For arrays this is last slice in view, for 3D this is depth-1, For remaining this is pitch-1
+    unsigned int DEPTH      : SQ_IMG_RSC_WRD4_DEPTH_SZ;
+    unsigned int            : 1; //Pitch[13] in gfx1030
+    unsigned int            : 2;
+    unsigned int BASE_ARRAY : SQ_IMG_RSC_WRD4_BASE_ARR_SZ;
+    unsigned int            : 3;
 #elif defined(BIGENDIAN_CPU)
-  unsigned int            : 3;
-  unsigned int BASE_ARRAY : SQ_IMG_RSC_WRD4_BASE_ARR_SZ;
-  unsigned int            : 2;
-  unsigned int            : 1; //Pitch[13] in gfx1030
-  unsigned int DEPTH      : SQ_IMG_RSC_WRD4_DEPTH_SZ; //Pitch[0:12] in gfx1030
+    unsigned int            : 3;
+    unsigned int BASE_ARRAY : SQ_IMG_RSC_WRD4_BASE_ARR_SZ;
+    unsigned int            : 2;
+    unsigned int            : 1; //Pitch[13] in gfx1030
+    unsigned int DEPTH      : SQ_IMG_RSC_WRD4_DEPTH_SZ; //Pitch[0:12] in gfx1030
 #endif
+  };
+  struct {
+#if defined(LITTLEENDIAN_CPU)
+    // For 1d, 2d and 2d-msaa in gfx1030 this is pitch-1
+    unsigned int PITCH      : SQ_IMG_RSC_WRD4_PITCH_SZ;
+    unsigned int            : SQ_IMG_RSC_WRD4_REG_SZ-SQ_IMG_RSC_WRD4_PITCH_SZ;
+#elif defined(BIGENDIAN_CPU)
+    unsigned int            : SQ_IMG_RSC_WRD4_REG_SZ-SQ_IMG_RSC_WRD4_PITCH_SZ;
+    unsigned int PITCH      : SQ_IMG_RSC_WRD4_PITCH_SZ;
+#endif
+  };
 };
 union SQ_IMG_RSRC_WORD4 {
   sq_img_rsrc_word4_t bitfields, bits, f;