gpu: preserve zero-stride buffer fetch payloads

tests/ShaderRecompilerComputeTests.cpp

@@ -10855,6 +10855,102 @@ public:
     std::printf("[gpu]     %-32s ok\n", name);
   }
 
+  void CheckZeroStrideBufferBinding() {
+    constexpr const char *name = "ZeroStrideBufferBinding";
+    constexpr uintptr_t base = 0x0000000204c00000ull;
+    constexpr uint64_t size = 0x10000;
+    constexpr auto output = base + 0x2000;
+    constexpr u32 value = 0x80000578u;
+    std::array<std::vector<u32>, 2> codes;
+    for (u32 scalar = 0; scalar < codes.size(); ++scalar) {
+      auto &code = codes[scalar];
+      if (scalar) {
+        code = {EncodeSmem0(0x08, 16), EncodeSmem1(0, 125),
+                EncodeSopp(0x0c, 0), EncodeVop1(0x01, 1, 16)};
+      } else {
+        code = {EncodeMubuf0(0x00, 0, false, false),
+                EncodeMubuf1(1, 0, 0), EncodeSopp(0x0c, 0)};
+      }
+      code.insert(code.end(), {EncodeVop2(0x1a, 2, InlineU32(2), 0),
+                              EncodeMubuf0(0x1c), EncodeMubuf1(1, 1, 2)});
+      AppendEnd(&code);
+      ShaderMapUserData(reinterpret_cast<uint64_t>(code.data()),
+          {.type = Prospero::ShaderBinaryType::kCs,
+           .code_size_bytes = static_cast<u32>(code.size() * sizeof(u32))});
+    }
+    EnsureRuntimeContext();
+    int64_t physical = -1;
+    Require(name, "allocate", LibKernel::Memory::KernelAllocateDirectMemory(
+        0, LibKernel::Memory::KernelGetDirectMemorySize(), size, size, 0, &physical) == 0,
+        "buffer binding allocation failed");
+    void *mapped = reinterpret_cast<void *>(base);
+    Require(name, "map", LibKernel::Memory::KernelMapDirectMemory(
+        &mapped, size, 0x3, 0x10, physical, size) == 0 &&
+            mapped == reinterpret_cast<void *>(base), "buffer binding mapping failed");
+    for (const auto prefix : {0u, 8u})
+      Require(name, "input backing", LibKernel::Memory::TryWriteBacking(
+          base + 0x1000 + prefix, &value, sizeof(value)), "input DWORD could not be written");
+    RenderContext context(m_runtime_context);
+    context.InitializeGpu(nullptr);
+    LibKernel::Memory::InstallGpuResources(&context);
+    context.GetGpu().SendCommandSync([&] {
+      CommandProcessor processor(context, 0);
+      processor.BufferInit();
+      context.MapMemory(base, size);
+      auto &cache = context.GetBufferCache();
+      auto &shaders = processor.GetShCtx();
+      struct Case { u32 mode, records; bool scalar; u32 expected; };
+      constexpr Case cases[]{{2, 1, false, value}, {2, 0, false, 0},
+                              {3, 1, false, 0}, {3, 4, false, value},
+                              {2, 1, true, 0}, {2, 4, true, value}};
+      for (const auto wave : {32u, 64u}) {
+        for (const auto prefix : {0u, 8u}) {
+          const auto address = base + 0x1000 + prefix;
+          for (const auto test : cases) {
+            cache.FillBuffer(output, wave * sizeof(u32), 0xa5a5a5a5u, false);
+            shaders.SetCsShader({
+                .data_addr = reinterpret_cast<uint64_t>(codes[test.scalar].data()),
+                .num_thread_x = wave, .num_thread_y = 1, .num_thread_z = 1,
+                .wave_size = static_cast<uint8_t>(wave), .user_sgpr = 8});
+            ShaderBufferResource input{}, out{};
+            input.UpdateAddress48(address);
+            input.fields[2] = test.records;
+            input.fields[3] = DstSel(4, 5, 6, 7) |
+                (static_cast<u32>(Prospero::BufferFormat::k32UInt) << 12u) |
+                (test.mode << 28u);
+            out.UpdateAddress48(output);
+            out.fields[2] = wave * sizeof(u32);
+            out.fields[3] = DstSel(4, 5, 6, 7) |
+                (static_cast<u32>(Prospero::BufferFormat::k32UInt) << 12u) | (3u << 28u);
+            for (u32 i = 0; i < 4; ++i) {
+              shaders.SetCsUserSgpr(i, input.fields[i], HW::UserSgprType::Unknown);
+              shaders.SetCsUserSgpr(i + 4, out.fields[i], HW::UserSgprType::Unknown);
+            }
+            processor.DispatchDirect(1, 1, 1, wave == 32 ? 0x8041u : 0x41u);
+            cache.ReadMemory(output, wave * sizeof(u32));
+            std::array<u32, 64> actual{};
+            Require(name, "readback", LibKernel::Memory::TryReadBacking(
+                output, actual.data(), wave * sizeof(u32)), "buffer output was unreadable");
+            Require(name, "vector and scalar range rules",
+                std::all_of(actual.begin(), actual.begin() + wave,
+                            [&](u32 word) { return word == test.expected; }),
+                "wave=" + std::to_string(wave) + " prefix=" + std::to_string(prefix) +
+                    " mode=" + std::to_string(test.mode) +
+                    " records=" + std::to_string(test.records) +
+                    " scalar=" + std::to_string(test.scalar));
+          }
+        }
+      }
+      context.UnmapMemory(base, size);
+      context.GetCommandScheduler().Finish();
+    });
+    LibKernel::Memory::InstallGpuResources(nullptr);
+    context.ShutdownGpu();
+    Require(name, "release", LibKernel::Memory::KernelCheckedReleaseDirectMemory(
+        physical, size) == 0, "buffer binding release failed");
+    std::printf("[gpu]     %-32s ok\n", name);
+  }
+
   void CheckOrderedDispatch() {
     constexpr const char *name = "OrderedDispatch";
     constexpr uintptr_t base = 0x0000000204a00000ull;
@@ -49546,6 +49642,11 @@ int main(int argc, char **argv) {
     vulkan.CheckOrderedDispatch();
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--zero-stride-binding-only") == 0) {
+    VulkanHarness vulkan;
+    vulkan.CheckZeroStrideBufferBinding();
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--layered-image-only") == 0) {
     VulkanHarness vulkan;
     vulkan.CheckUnifiedImageViewCache();
@@ -49799,6 +49900,7 @@ int main(int argc, char **argv) {
   vulkan.CheckNativeIndirectDispatch();
   CheckComputeLdsLimit(vulkan);
   vulkan.CheckOrderedDispatch();
+  vulkan.CheckZeroStrideBufferBinding();
   vulkan.CheckNullTextureDescriptors();
   vulkan.CheckBdaPageTableUploads();
   vulkan.CheckBdaSyncUploads();