Merge pull request #10996 from Kelebek1/readblock_optimisation

Use spans over guest memory where possible instead of copying data

12 files changed, 144 insertions, 139 deletions

diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index b5ed3380f..6ed4b78f2 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -234,9 +234,10 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
     if (has_new_downloads) {
         memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
     }
-    tmp_buffer.resize_destructive(amount);
-    cpu_memory.ReadBlockUnsafe(*cpu_src_address, tmp_buffer.data(), amount);
-    cpu_memory.WriteBlockUnsafe(*cpu_dest_address, tmp_buffer.data(), amount);
+
+    Core::Memory::CpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::UnsafeReadWrite> tmp(
+        cpu_memory, *cpu_src_address, amount, &tmp_buffer);
+    tmp.SetAddressAndSize(*cpu_dest_address, amount);
     return true;
 }
 
diff --git a/src/video_core/dma_pusher.cpp b/src/video_core/dma_pusher.cpp
index 551929824..9f1b340a9 100644
--- a/src/video_core/dma_pusher.cpp
+++ b/src/video_core/dma_pusher.cpp
@@ -5,6 +5,7 @@
 #include "common/microprofile.h"
 #include "common/settings.h"
 #include "core/core.h"
+#include "core/memory.h"
 #include "video_core/dma_pusher.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/gpu.h"
@@ -12,6 +13,8 @@
 
 namespace Tegra {
 
+constexpr u32 MacroRegistersStart = 0xE00;
+
 DmaPusher::DmaPusher(Core::System& system_, GPU& gpu_, MemoryManager& memory_manager_,
                      Control::ChannelState& channel_state_)
     : gpu{gpu_}, system{system_}, memory_manager{memory_manager_}, puller{gpu_, memory_manager_,
@@ -74,25 +77,16 @@ bool DmaPusher::Step() {
         }
 
         // Push buffer non-empty, read a word
-        command_headers.resize_destructive(command_list_header.size);
-        constexpr u32 MacroRegistersStart = 0xE00;
-        if (dma_state.method < MacroRegistersStart) {
-            if (Settings::IsGPULevelHigh()) {
-                memory_manager.ReadBlock(dma_state.dma_get, command_headers.data(),
-                                         command_list_header.size * sizeof(u32));
-            } else {
-                memory_manager.ReadBlockUnsafe(dma_state.dma_get, command_headers.data(),
-                                               command_list_header.size * sizeof(u32));
-            }
-        } else {
-            const size_t copy_size = command_list_header.size * sizeof(u32);
+        if (dma_state.method >= MacroRegistersStart) {
             if (subchannels[dma_state.subchannel]) {
-                subchannels[dma_state.subchannel]->current_dirty =
-                    memory_manager.IsMemoryDirty(dma_state.dma_get, copy_size);
+                subchannels[dma_state.subchannel]->current_dirty = memory_manager.IsMemoryDirty(
+                    dma_state.dma_get, command_list_header.size * sizeof(u32));
             }
-            memory_manager.ReadBlockUnsafe(dma_state.dma_get, command_headers.data(), copy_size);
         }
-        ProcessCommands(command_headers);
+        Core::Memory::GpuGuestMemory<Tegra::CommandHeader,
+                                     Core::Memory::GuestMemoryFlags::UnsafeRead>
+            headers(memory_manager, dma_state.dma_get, command_list_header.size, &command_headers);
+        ProcessCommands(headers);
     }
 
     return true;
diff --git a/src/video_core/engines/engine_upload.cpp b/src/video_core/engines/engine_upload.cpp
index 7f5a0c29d..bc64d4486 100644
--- a/src/video_core/engines/engine_upload.cpp
+++ b/src/video_core/engines/engine_upload.cpp
@@ -5,6 +5,7 @@
 
 #include "common/algorithm.h"
 #include "common/assert.h"
+#include "core/memory.h"
 #include "video_core/engines/engine_upload.h"
 #include "video_core/memory_manager.h"
 #include "video_core/rasterizer_interface.h"
@@ -46,15 +47,11 @@ void State::ProcessData(const u32* data, size_t num_data) {
 void State::ProcessData(std::span<const u8> read_buffer) {
     const GPUVAddr address{regs.dest.Address()};
     if (is_linear) {
-        if (regs.line_count == 1) {
-            rasterizer->AccelerateInlineToMemory(address, copy_size, read_buffer);
-        } else {
-            for (size_t line = 0; line < regs.line_count; ++line) {
-                const GPUVAddr dest_line = address + line * regs.dest.pitch;
-                std::span<const u8> buffer(read_buffer.data() + line * regs.line_length_in,
-                                           regs.line_length_in);
-                rasterizer->AccelerateInlineToMemory(dest_line, regs.line_length_in, buffer);
-            }
+        for (size_t line = 0; line < regs.line_count; ++line) {
+            const GPUVAddr dest_line = address + line * regs.dest.pitch;
+            std::span<const u8> buffer(read_buffer.data() + line * regs.line_length_in,
+                                       regs.line_length_in);
+            rasterizer->AccelerateInlineToMemory(dest_line, regs.line_length_in, buffer);
         }
     } else {
         u32 width = regs.dest.width;
@@ -70,13 +67,14 @@ void State::ProcessData(std::span<const u8> read_buffer) {
         const std::size_t dst_size = Tegra::Texture::CalculateSize(
             true, bytes_per_pixel, width, regs.dest.height, regs.dest.depth,
             regs.dest.BlockHeight(), regs.dest.BlockDepth());
-        tmp_buffer.resize_destructive(dst_size);
-        memory_manager.ReadBlock(address, tmp_buffer.data(), dst_size);
-        Tegra::Texture::SwizzleSubrect(tmp_buffer, read_buffer, bytes_per_pixel, width,
-                                       regs.dest.height, regs.dest.depth, x_offset, regs.dest.y,
-                                       x_elements, regs.line_count, regs.dest.BlockHeight(),
+
+        Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+            tmp(memory_manager, address, dst_size, &tmp_buffer);
+
+        Tegra::Texture::SwizzleSubrect(tmp, read_buffer, bytes_per_pixel, width, regs.dest.height,
+                                       regs.dest.depth, x_offset, regs.dest.y, x_elements,
+                                       regs.line_count, regs.dest.BlockHeight(),
                                        regs.dest.BlockDepth(), regs.line_length_in);
-        memory_manager.WriteBlockCached(address, tmp_buffer.data(), dst_size);
     }
 }
 
diff --git a/src/video_core/engines/kepler_compute.cpp b/src/video_core/engines/kepler_compute.cpp
index 601095f03..a38d9528a 100644
--- a/src/video_core/engines/kepler_compute.cpp
+++ b/src/video_core/engines/kepler_compute.cpp
@@ -84,7 +84,6 @@ Texture::TICEntry KeplerCompute::GetTICEntry(u32 tic_index) const {
 
     Texture::TICEntry tic_entry;
     memory_manager.ReadBlockUnsafe(tic_address_gpu, &tic_entry, sizeof(Texture::TICEntry));
-
     return tic_entry;
 }
 
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 62d70e9f3..c3696096d 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -9,6 +9,7 @@
 #include "common/settings.h"
 #include "core/core.h"
 #include "core/core_timing.h"
+#include "core/memory.h"
 #include "video_core/dirty_flags.h"
 #include "video_core/engines/draw_manager.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -679,17 +680,14 @@ void Maxwell3D::ProcessCBData(u32 value) {
 Texture::TICEntry Maxwell3D::GetTICEntry(u32 tic_index) const {
     const GPUVAddr tic_address_gpu{regs.tex_header.Address() +
                                    tic_index * sizeof(Texture::TICEntry)};
-
     Texture::TICEntry tic_entry;
     memory_manager.ReadBlockUnsafe(tic_address_gpu, &tic_entry, sizeof(Texture::TICEntry));
-
     return tic_entry;
 }
 
 Texture::TSCEntry Maxwell3D::GetTSCEntry(u32 tsc_index) const {
     const GPUVAddr tsc_address_gpu{regs.tex_sampler.Address() +
                                    tsc_index * sizeof(Texture::TSCEntry)};
-
     Texture::TSCEntry tsc_entry;
     memory_manager.ReadBlockUnsafe(tsc_address_gpu, &tsc_entry, sizeof(Texture::TSCEntry));
     return tsc_entry;
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index f8598fd98..cd8e24b0b 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -7,6 +7,7 @@
 #include "common/microprofile.h"
 #include "common/settings.h"
 #include "core/core.h"
+#include "core/memory.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/memory_manager.h"
@@ -130,11 +131,12 @@ void MaxwellDMA::Launch() {
                 UNIMPLEMENTED_IF(regs.offset_out % 16 != 0);
                 read_buffer.resize_destructive(16);
                 for (u32 offset = 0; offset < regs.line_length_in; offset += 16) {
-                    memory_manager.ReadBlock(
-                        convert_linear_2_blocklinear_addr(regs.offset_in + offset),
-                        read_buffer.data(), read_buffer.size());
-                    memory_manager.WriteBlockCached(regs.offset_out + offset, read_buffer.data(),
-                                                    read_buffer.size());
+                    Core::Memory::GpuGuestMemoryScoped<
+                        u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+                        tmp_write_buffer(memory_manager,
+                                         convert_linear_2_blocklinear_addr(regs.offset_in + offset),
+                                         16, &read_buffer);
+                    tmp_write_buffer.SetAddressAndSize(regs.offset_out + offset, 16);
                 }
             } else if (is_src_pitch && !is_dst_pitch) {
                 UNIMPLEMENTED_IF(regs.line_length_in % 16 != 0);
@@ -142,20 +144,19 @@ void MaxwellDMA::Launch() {
                 UNIMPLEMENTED_IF(regs.offset_out % 16 != 0);
                 read_buffer.resize_destructive(16);
                 for (u32 offset = 0; offset < regs.line_length_in; offset += 16) {
-                    memory_manager.ReadBlock(regs.offset_in + offset, read_buffer.data(),
-                                             read_buffer.size());
-                    memory_manager.WriteBlockCached(
-                        convert_linear_2_blocklinear_addr(regs.offset_out + offset),
-                        read_buffer.data(), read_buffer.size());
+                    Core::Memory::GpuGuestMemoryScoped<
+                        u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+                        tmp_write_buffer(memory_manager, regs.offset_in + offset, 16, &read_buffer);
+                    tmp_write_buffer.SetAddressAndSize(
+                        convert_linear_2_blocklinear_addr(regs.offset_out + offset), 16);
                 }
             } else {
                 if (!accelerate.BufferCopy(regs.offset_in, regs.offset_out, regs.line_length_in)) {
-                    read_buffer.resize_destructive(regs.line_length_in);
-                    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(),
-                                             regs.line_length_in,
-                                             VideoCommon::CacheType::NoBufferCache);
-                    memory_manager.WriteBlockCached(regs.offset_out, read_buffer.data(),
-                                                    regs.line_length_in);
+                    Core::Memory::GpuGuestMemoryScoped<
+                        u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+                        tmp_write_buffer(memory_manager, regs.offset_in, regs.line_length_in,
+                                         &read_buffer);
+                    tmp_write_buffer.SetAddressAndSize(regs.offset_out, regs.line_length_in);
                 }
             }
         }
@@ -222,17 +223,15 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
         CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
 
     const size_t dst_size = dst_operand.pitch * regs.line_count;
-    read_buffer.resize_destructive(src_size);
-    write_buffer.resize_destructive(dst_size);
 
-    memory_manager.ReadBlock(src_operand.address, read_buffer.data(), src_size);
-    memory_manager.ReadBlock(dst_operand.address, write_buffer.data(), dst_size);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_read_buffer(
+        memory_manager, src_operand.address, src_size, &read_buffer);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+        tmp_write_buffer(memory_manager, dst_operand.address, dst_size, &write_buffer);
 
-    UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
-                     src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
-                     dst_operand.pitch);
-
-    memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
+    UnswizzleSubrect(tmp_write_buffer, tmp_read_buffer, bytes_per_pixel, width, height, depth,
+                     x_offset, src_params.origin.y, x_elements, regs.line_count, block_height,
+                     block_depth, dst_operand.pitch);
 }
 
 void MaxwellDMA::CopyPitchToBlockLinear() {
@@ -287,18 +286,17 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
         CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
     const size_t src_size = static_cast<size_t>(regs.pitch_in) * regs.line_count;
 
-    read_buffer.resize_destructive(src_size);
-    write_buffer.resize_destructive(dst_size);
-
-    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
-    memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
-
-    // If the input is linear and the output is tiled, swizzle the input and copy it over.
-    SwizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
-                   dst_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
-                   regs.pitch_in);
-
-    memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
+    GPUVAddr src_addr = regs.offset_in;
+    GPUVAddr dst_addr = regs.offset_out;
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_read_buffer(
+        memory_manager, src_addr, src_size, &read_buffer);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+        tmp_write_buffer(memory_manager, dst_addr, dst_size, &write_buffer);
+
+    //  If the input is linear and the output is tiled, swizzle the input and copy it over.
+    SwizzleSubrect(tmp_write_buffer, tmp_read_buffer, bytes_per_pixel, width, height, depth,
+                   x_offset, dst_params.origin.y, x_elements, regs.line_count, block_height,
+                   block_depth, regs.pitch_in);
 }
 
 void MaxwellDMA::CopyBlockLinearToBlockLinear() {
@@ -342,23 +340,20 @@ void MaxwellDMA::CopyBlockLinearToBlockLinear() {
     const u32 pitch = x_elements * bytes_per_pixel;
     const size_t mid_buffer_size = pitch * regs.line_count;
 
-    read_buffer.resize_destructive(src_size);
-    write_buffer.resize_destructive(dst_size);
-
     intermediate_buffer.resize_destructive(mid_buffer_size);
 
-    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
-    memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_read_buffer(
+        memory_manager, regs.offset_in, src_size, &read_buffer);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadCachedWrite>
+        tmp_write_buffer(memory_manager, regs.offset_out, dst_size, &write_buffer);
 
-    UnswizzleSubrect(intermediate_buffer, read_buffer, bytes_per_pixel, src_width, src.height,
+    UnswizzleSubrect(intermediate_buffer, tmp_read_buffer, bytes_per_pixel, src_width, src.height,
                      src.depth, src_x_offset, src.origin.y, x_elements, regs.line_count,
                      src.block_size.height, src.block_size.depth, pitch);
 
-    SwizzleSubrect(write_buffer, intermediate_buffer, bytes_per_pixel, dst_width, dst.height,
+    SwizzleSubrect(tmp_write_buffer, intermediate_buffer, bytes_per_pixel, dst_width, dst.height,
                    dst.depth, dst_x_offset, dst.origin.y, x_elements, regs.line_count,
                    dst.block_size.height, dst.block_size.depth, pitch);
-
-    memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
 }
 
 void MaxwellDMA::ReleaseSemaphore() {
diff --git a/src/video_core/engines/sw_blitter/blitter.cpp b/src/video_core/engines/sw_blitter/blitter.cpp
index ff88cd03d..3a599f466 100644
--- a/src/video_core/engines/sw_blitter/blitter.cpp
+++ b/src/video_core/engines/sw_blitter/blitter.cpp
@@ -159,11 +159,11 @@ bool SoftwareBlitEngine::Blit(Fermi2D::Surface& src, Fermi2D::Surface& dst,
     const auto src_bytes_per_pixel = BytesPerBlock(PixelFormatFromRenderTargetFormat(src.format));
     const auto dst_bytes_per_pixel = BytesPerBlock(PixelFormatFromRenderTargetFormat(dst.format));
     const size_t src_size = get_surface_size(src, src_bytes_per_pixel);
-    impl->tmp_buffer.resize_destructive(src_size);
-    memory_manager.ReadBlock(src.Address(), impl->tmp_buffer.data(), src_size);
 
-    const size_t src_copy_size = src_extent_x * src_extent_y * src_bytes_per_pixel;
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::SafeRead> tmp_buffer(
+        memory_manager, src.Address(), src_size, &impl->tmp_buffer);
 
+    const size_t src_copy_size = src_extent_x * src_extent_y * src_bytes_per_pixel;
     const size_t dst_copy_size = dst_extent_x * dst_extent_y * dst_bytes_per_pixel;
 
     impl->src_buffer.resize_destructive(src_copy_size);
@@ -200,12 +200,11 @@ bool SoftwareBlitEngine::Blit(Fermi2D::Surface& src, Fermi2D::Surface& dst,
 
     impl->dst_buffer.resize_destructive(dst_copy_size);
     if (src.linear == Fermi2D::MemoryLayout::BlockLinear) {
-        UnswizzleSubrect(impl->src_buffer, impl->tmp_buffer, src_bytes_per_pixel, src.width,
-                         src.height, src.depth, config.src_x0, config.src_y0, src_extent_x,
-                         src_extent_y, src.block_height, src.block_depth,
-                         src_extent_x * src_bytes_per_pixel);
+        UnswizzleSubrect(impl->src_buffer, tmp_buffer, src_bytes_per_pixel, src.width, src.height,
+                         src.depth, config.src_x0, config.src_y0, src_extent_x, src_extent_y,
+                         src.block_height, src.block_depth, src_extent_x * src_bytes_per_pixel);
     } else {
-        process_pitch_linear(false, impl->tmp_buffer, impl->src_buffer, src_extent_x, src_extent_y,
+        process_pitch_linear(false, tmp_buffer, impl->src_buffer, src_extent_x, src_extent_y,
                              src.pitch, config.src_x0, config.src_y0, src_bytes_per_pixel);
     }
 
@@ -221,20 +220,18 @@ bool SoftwareBlitEngine::Blit(Fermi2D::Surface& src, Fermi2D::Surface& dst,
     }
 
     const size_t dst_size = get_surface_size(dst, dst_bytes_per_pixel);
-    impl->tmp_buffer.resize_destructive(dst_size);
-    memory_manager.ReadBlock(dst.Address(), impl->tmp_buffer.data(), dst_size);
+    Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::SafeReadWrite>
+        tmp_buffer2(memory_manager, dst.Address(), dst_size, &impl->tmp_buffer);
 
     if (dst.linear == Fermi2D::MemoryLayout::BlockLinear) {
-        SwizzleSubrect(impl->tmp_buffer, impl->dst_buffer, dst_bytes_per_pixel, dst.width,
-                       dst.height, dst.depth, config.dst_x0, config.dst_y0, dst_extent_x,
-                       dst_extent_y, dst.block_height, dst.block_depth,
-                       dst_extent_x * dst_bytes_per_pixel);
+        SwizzleSubrect(tmp_buffer2, impl->dst_buffer, dst_bytes_per_pixel, dst.width, dst.height,
+                       dst.depth, config.dst_x0, config.dst_y0, dst_extent_x, dst_extent_y,
+                       dst.block_height, dst.block_depth, dst_extent_x * dst_bytes_per_pixel);
     } else {
-        process_pitch_linear(true, impl->dst_buffer, impl->tmp_buffer, dst_extent_x, dst_extent_y,
+        process_pitch_linear(true, impl->dst_buffer, tmp_buffer2, dst_extent_x, dst_extent_y,
                              dst.pitch, config.dst_x0, config.dst_y0,
                              static_cast<size_t>(dst_bytes_per_pixel));
     }
-    memory_manager.WriteBlock(dst.Address(), impl->tmp_buffer.data(), dst_size);
     return true;
 }
 
diff --git a/src/video_core/memory_manager.cpp b/src/video_core/memory_manager.cpp
index 45141e488..d16040613 100644
--- a/src/video_core/memory_manager.cpp
+++ b/src/video_core/memory_manager.cpp
@@ -10,13 +10,13 @@
 #include "core/device_memory.h"
 #include "core/hle/kernel/k_page_table.h"
 #include "core/hle/kernel/k_process.h"
-#include "core/memory.h"
 #include "video_core/invalidation_accumulator.h"
 #include "video_core/memory_manager.h"
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_base.h"
 
 namespace Tegra {
+using Core::Memory::GuestMemoryFlags;
 
 std::atomic<size_t> MemoryManager::unique_identifier_generator{};
 
@@ -587,13 +587,10 @@ void MemoryManager::InvalidateRegion(GPUVAddr gpu_addr, size_t size,
 
 void MemoryManager::CopyBlock(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr, std::size_t size,
                               VideoCommon::CacheType which) {
-    tmp_buffer.resize_destructive(size);
-    ReadBlock(gpu_src_addr, tmp_buffer.data(), size, which);
-
-    // The output block must be flushed in case it has data modified from the GPU.
-    // Fixes NPC geometry in Zombie Panic in Wonderland DX
+    Core::Memory::GpuGuestMemoryScoped<u8, GuestMemoryFlags::SafeReadWrite> data(
+        *this, gpu_src_addr, size);
+    data.SetAddressAndSize(gpu_dest_addr, size);
     FlushRegion(gpu_dest_addr, size, which);
-    WriteBlock(gpu_dest_addr, tmp_buffer.data(), size, which);
 }
 
 bool MemoryManager::IsGranularRange(GPUVAddr gpu_addr, std::size_t size) const {
@@ -758,4 +755,23 @@ void MemoryManager::FlushCaching() {
     accumulator->Clear();
 }
 
+const u8* MemoryManager::GetSpan(const GPUVAddr src_addr, const std::size_t size) const {
+    auto cpu_addr = GpuToCpuAddress(src_addr);
+    if (cpu_addr) {
+        return memory.GetSpan(*cpu_addr, size);
+    }
+    return nullptr;
+}
+
+u8* MemoryManager::GetSpan(const GPUVAddr src_addr, const std::size_t size) {
+    if (!IsContinuousRange(src_addr, size)) {
+        return nullptr;
+    }
+    auto cpu_addr = GpuToCpuAddress(src_addr);
+    if (cpu_addr) {
+        return memory.GetSpan(*cpu_addr, size);
+    }
+    return nullptr;
+}
+
 } // namespace Tegra
diff --git a/src/video_core/memory_manager.h b/src/video_core/memory_manager.h
index 4202c26ff..9b311b9e5 100644
--- a/src/video_core/memory_manager.h
+++ b/src/video_core/memory_manager.h
@@ -15,6 +15,7 @@
 #include "common/range_map.h"
 #include "common/scratch_buffer.h"
 #include "common/virtual_buffer.h"
+#include "core/memory.h"
 #include "video_core/cache_types.h"
 #include "video_core/pte_kind.h"
 
@@ -62,6 +63,20 @@ public:
     [[nodiscard]] u8* GetPointer(GPUVAddr addr);
     [[nodiscard]] const u8* GetPointer(GPUVAddr addr) const;
 
+    template <typename T>
+    [[nodiscard]] T* GetPointer(GPUVAddr addr) {
+        const auto address{GpuToCpuAddress(addr)};
+        if (!address) {
+            return {};
+        }
+        return memory.GetPointer(*address);
+    }
+
+    template <typename T>
+    [[nodiscard]] const T* GetPointer(GPUVAddr addr) const {
+        return GetPointer<T*>(addr);
+    }
+
     /**
      * ReadBlock and WriteBlock are full read and write operations over virtual
      * GPU Memory. It's important to use these when GPU memory may not be continuous
@@ -139,6 +154,9 @@ public:
 
     void FlushCaching();
 
+    const u8* GetSpan(const GPUVAddr src_addr, const std::size_t size) const;
+    u8* GetSpan(const GPUVAddr src_addr, const std::size_t size);
+
 private:
     template <bool is_big_pages, typename FuncMapped, typename FuncReserved, typename FuncUnmapped>
     inline void MemoryOperation(GPUVAddr gpu_src_addr, std::size_t size, FuncMapped&& func_mapped,
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 3a859139c..4457b366f 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -8,6 +8,7 @@
 
 #include "common/alignment.h"
 #include "common/settings.h"
+#include "core/memory.h"
 #include "video_core/control/channel_state.h"
 #include "video_core/dirty_flags.h"
 #include "video_core/engines/kepler_compute.h"
@@ -1026,19 +1027,19 @@ void TextureCache<P>::UploadImageContents(Image& image, StagingBuffer& staging)
         runtime.AccelerateImageUpload(image, staging, uploads);
         return;
     }
-    const size_t guest_size_bytes = image.guest_size_bytes;
-    swizzle_data_buffer.resize_destructive(guest_size_bytes);
-    gpu_memory->ReadBlockUnsafe(gpu_addr, swizzle_data_buffer.data(), guest_size_bytes);
+
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> swizzle_data(
+        *gpu_memory, gpu_addr, image.guest_size_bytes, &swizzle_data_buffer);
 
     if (True(image.flags & ImageFlagBits::Converted)) {
         unswizzle_data_buffer.resize_destructive(image.unswizzled_size_bytes);
-        auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer,
-                                     unswizzle_data_buffer);
+        auto copies =
+            UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data, unswizzle_data_buffer);
         ConvertImage(unswizzle_data_buffer, image.info, mapped_span, copies);
         image.UploadMemory(staging, copies);
     } else {
         const auto copies =
-            UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer, mapped_span);
+            UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data, mapped_span);
         image.UploadMemory(staging, copies);
     }
 }
@@ -1231,11 +1232,12 @@ void TextureCache<P>::QueueAsyncDecode(Image& image, ImageId image_id) {
     decode->image_id = image_id;
     async_decodes.push_back(std::move(decode));
 
-    Common::ScratchBuffer<u8> local_unswizzle_data_buffer(image.unswizzled_size_bytes);
-    const size_t guest_size_bytes = image.guest_size_bytes;
-    swizzle_data_buffer.resize_destructive(guest_size_bytes);
-    gpu_memory->ReadBlockUnsafe(image.gpu_addr, swizzle_data_buffer.data(), guest_size_bytes);
-    auto copies = UnswizzleImage(*gpu_memory, image.gpu_addr, image.info, swizzle_data_buffer,
+    static Common::ScratchBuffer<u8> local_unswizzle_data_buffer;
+    local_unswizzle_data_buffer.resize_destructive(image.unswizzled_size_bytes);
+    Core::Memory::GpuGuestMemory<u8, Core::Memory::GuestMemoryFlags::UnsafeRead> swizzle_data(
+        *gpu_memory, image.gpu_addr, image.guest_size_bytes, &swizzle_data_buffer);
+
+    auto copies = UnswizzleImage(*gpu_memory, image.gpu_addr, image.info, swizzle_data,
                                  local_unswizzle_data_buffer);
     const size_t out_size = MapSizeBytes(image);
 
diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp
index 0de6ed09d..a83f5d41c 100644
--- a/src/video_core/texture_cache/util.cpp
+++ b/src/video_core/texture_cache/util.cpp
@@ -20,6 +20,7 @@
 #include "common/div_ceil.h"
 #include "common/scratch_buffer.h"
 #include "common/settings.h"
+#include "core/memory.h"
 #include "video_core/compatible_formats.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/memory_manager.h"
@@ -544,17 +545,15 @@ void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr
                        tile_size.height, info.tile_width_spacing);
     const size_t subresource_size = sizes[level];
 
-    tmp_buffer.resize_destructive(subresource_size);
-    const std::span<u8> dst(tmp_buffer);
-
     for (s32 layer = 0; layer < info.resources.layers; ++layer) {
         const std::span<const u8> src = input.subspan(host_offset);
-        gpu_memory.ReadBlockUnsafe(gpu_addr + guest_offset, dst.data(), dst.size_bytes());
-
-        SwizzleTexture(dst, src, bytes_per_block, num_tiles.width, num_tiles.height,
-                       num_tiles.depth, block.height, block.depth);
+        {
+            Core::Memory::GpuGuestMemoryScoped<u8, Core::Memory::GuestMemoryFlags::UnsafeReadWrite>
+                dst(gpu_memory, gpu_addr + guest_offset, subresource_size, &tmp_buffer);
 
-        gpu_memory.WriteBlockUnsafe(gpu_addr + guest_offset, dst.data(), dst.size_bytes());
+            SwizzleTexture(dst, src, bytes_per_block, num_tiles.width, num_tiles.height,
+                           num_tiles.depth, block.height, block.depth);
+        }
 
         host_offset += host_bytes_per_layer;
         guest_offset += layer_stride;
@@ -837,6 +836,7 @@ boost::container::small_vector<BufferImageCopy, 16> UnswizzleImage(Tegra::Memory
     const Extent3D size = info.size;
 
     if (info.type == ImageType::Linear) {
+        ASSERT(output.size_bytes() >= guest_size_bytes);
         gpu_memory.ReadBlockUnsafe(gpu_addr, output.data(), guest_size_bytes);
 
         ASSERT((info.pitch >> bpp_log2) << bpp_log2 == info.pitch);
@@ -904,16 +904,6 @@ boost::container::small_vector<BufferImageCopy, 16> UnswizzleImage(Tegra::Memory
     return copies;
 }
 
-BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
-                            const ImageBase& image, std::span<u8> output) {
-    gpu_memory.ReadBlockUnsafe(gpu_addr, output.data(), image.guest_size_bytes);
-    return BufferCopy{
-        .src_offset = 0,
-        .dst_offset = 0,
-        .size = image.guest_size_bytes,
-    };
-}
-
 void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
                   std::span<BufferImageCopy> copies) {
     u32 output_offset = 0;
diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h
index ab45a43c4..5a0649d24 100644
--- a/src/video_core/texture_cache/util.h
+++ b/src/video_core/texture_cache/util.h
@@ -66,9 +66,6 @@ struct OverlapResult {
     Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info,
     std::span<const u8> input, std::span<u8> output);
 
-[[nodiscard]] BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr,
-                                          const ImageBase& image, std::span<u8> output);
-
 void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
                   std::span<BufferImageCopy> copies);