From 430255caf86afdce8475898d5a34dce38f955ecf Mon Sep 17 00:00:00 2001
From: yzct12345 <87620833+yzct12345@users.noreply.github.com>
Date: Thu, 12 Aug 2021 04:45:25 +0000
Subject: [PATCH] decoders: Templates allow memcpy optimizations

---
 src/video_core/textures/decoders.cpp | 235 +++++++++++++++++----------
 1 file changed, 147 insertions(+), 88 deletions(-)

diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index c32ae956a3..c010b9353a 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -84,34 +84,107 @@ template <bool TO_LINEAR>
 void Swizzle(std::span<u8> output, std::span<const u8> input, u32 bytes_per_pixel, u32 width,
              u32 height, u32 depth, u32 block_height, u32 block_depth, u32 stride_alignment) {
     switch (bytes_per_pixel) {
-    case 1:
-        return SwizzleImpl<TO_LINEAR, 1>(output, input, width, height, depth, block_height,
+#define BPP_CASE(x)                                                                                \
+    case x:                                                                                        \
+        return SwizzleImpl<TO_LINEAR, x>(output, input, width, height, depth, block_height,        \
                                          block_depth, stride_alignment);
-    case 2:
-        return SwizzleImpl<TO_LINEAR, 2>(output, input, width, height, depth, block_height,
-                                         block_depth, stride_alignment);
-    case 3:
-        return SwizzleImpl<TO_LINEAR, 3>(output, input, width, height, depth, block_height,
-                                         block_depth, stride_alignment);
-    case 4:
-        return SwizzleImpl<TO_LINEAR, 4>(output, input, width, height, depth, block_height,
-                                         block_depth, stride_alignment);
-    case 6:
-        return SwizzleImpl<TO_LINEAR, 6>(output, input, width, height, depth, block_height,
-                                         block_depth, stride_alignment);
-    case 8:
-        return SwizzleImpl<TO_LINEAR, 8>(output, input, width, height, depth, block_height,
-                                         block_depth, stride_alignment);
-    case 12:
-        return SwizzleImpl<TO_LINEAR, 12>(output, input, width, height, depth, block_height,
-                                          block_depth, stride_alignment);
-    case 16:
-        return SwizzleImpl<TO_LINEAR, 16>(output, input, width, height, depth, block_height,
-                                          block_depth, stride_alignment);
+        BPP_CASE(1)
+        BPP_CASE(2)
+        BPP_CASE(3)
+        BPP_CASE(4)
+        BPP_CASE(6)
+        BPP_CASE(8)
+        BPP_CASE(12)
+        BPP_CASE(16)
+#undef BPP_CASE
     default:
         UNREACHABLE_MSG("Invalid bytes_per_pixel={}", bytes_per_pixel);
     }
 }
+
+template <u32 BYTES_PER_PIXEL>
+void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width,
+                    u8* swizzled_data, const u8* unswizzled_data, u32 block_height_bit,
+                    u32 offset_x, u32 offset_y) {
+    const u32 block_height = 1U << block_height_bit;
+    const u32 image_width_in_gobs =
+        (swizzled_width * BYTES_PER_PIXEL + (GOB_SIZE_X - 1)) / GOB_SIZE_X;
+    for (u32 line = 0; line < subrect_height; ++line) {
+        const u32 dst_y = line + offset_y;
+        const u32 gob_address_y =
+            (dst_y / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height * image_width_in_gobs +
+            ((dst_y % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
+        const auto& table = SWIZZLE_TABLE[dst_y % GOB_SIZE_Y];
+        for (u32 x = 0; x < subrect_width; ++x) {
+            const u32 dst_x = x + offset_x;
+            const u32 gob_address =
+                gob_address_y + (dst_x * BYTES_PER_PIXEL / GOB_SIZE_X) * GOB_SIZE * block_height;
+            const u32 swizzled_offset = gob_address + table[(dst_x * BYTES_PER_PIXEL) % GOB_SIZE_X];
+            const u32 unswizzled_offset = line * source_pitch + x * BYTES_PER_PIXEL;
+
+            const u8* const source_line = unswizzled_data + unswizzled_offset;
+            u8* const dest_addr = swizzled_data + swizzled_offset;
+            std::memcpy(dest_addr, source_line, BYTES_PER_PIXEL);
+        }
+    }
+}
+
+template <u32 BYTES_PER_PIXEL>
+void UnswizzleSubrect(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 block_height,
+                      u32 origin_x, u32 origin_y, u8* output, const u8* input) {
+    const u32 stride = width * BYTES_PER_PIXEL;
+    const u32 gobs_in_x = (stride + GOB_SIZE_X - 1) / GOB_SIZE_X;
+    const u32 block_size = gobs_in_x << (GOB_SIZE_SHIFT + block_height);
+
+    const u32 block_height_mask = (1U << block_height) - 1;
+    const u32 x_shift = GOB_SIZE_SHIFT + block_height;
+
+    for (u32 line = 0; line < line_count; ++line) {
+        const u32 src_y = line + origin_y;
+        const auto& table = SWIZZLE_TABLE[src_y % GOB_SIZE_Y];
+
+        const u32 block_y = src_y >> GOB_SIZE_Y_SHIFT;
+        const u32 src_offset_y = (block_y >> block_height) * block_size +
+                                 ((block_y & block_height_mask) << GOB_SIZE_SHIFT);
+        for (u32 column = 0; column < line_length_in; ++column) {
+            const u32 src_x = (column + origin_x) * BYTES_PER_PIXEL;
+            const u32 src_offset_x = (src_x >> GOB_SIZE_X_SHIFT) << x_shift;
+
+            const u32 swizzled_offset = src_offset_y + src_offset_x + table[src_x % GOB_SIZE_X];
+            const u32 unswizzled_offset = line * pitch + column * BYTES_PER_PIXEL;
+
+            std::memcpy(output + unswizzled_offset, input + swizzled_offset, BYTES_PER_PIXEL);
+        }
+    }
+}
+
+template <u32 BYTES_PER_PIXEL>
+void SwizzleSliceToVoxel(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 height,
+                         u32 block_height, u32 block_depth, u32 origin_x, u32 origin_y, u8* output,
+                         const u8* input) {
+    UNIMPLEMENTED_IF(origin_x > 0);
+    UNIMPLEMENTED_IF(origin_y > 0);
+
+    const u32 stride = width * BYTES_PER_PIXEL;
+    const u32 gobs_in_x = (stride + GOB_SIZE_X - 1) / GOB_SIZE_X;
+    const u32 block_size = gobs_in_x << (GOB_SIZE_SHIFT + block_height + block_depth);
+
+    const u32 block_height_mask = (1U << block_height) - 1;
+    const u32 x_shift = static_cast<u32>(GOB_SIZE_SHIFT) + block_height + block_depth;
+
+    for (u32 line = 0; line < line_count; ++line) {
+        const auto& table = SWIZZLE_TABLE[line % GOB_SIZE_Y];
+        const u32 block_y = line / GOB_SIZE_Y;
+        const u32 dst_offset_y =
+            (block_y >> block_height) * block_size + (block_y & block_height_mask) * GOB_SIZE;
+        for (u32 x = 0; x < line_length_in; ++x) {
+            const u32 dst_offset =
+                ((x / GOB_SIZE_X) << x_shift) + dst_offset_y + table[x % GOB_SIZE_X];
+            const u32 src_offset = x * BYTES_PER_PIXEL + line * pitch;
+            std::memcpy(output + dst_offset, input + src_offset, BYTES_PER_PIXEL);
+        }
+    }
+}
 } // Anonymous namespace
 
 void UnswizzleTexture(std::span<u8> output, std::span<const u8> input, u32 bytes_per_pixel,
@@ -131,81 +204,67 @@ void SwizzleTexture(std::span<u8> output, std::span<const u8> input, u32 bytes_p
 void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width,
                     u32 bytes_per_pixel, u8* swizzled_data, const u8* unswizzled_data,
                     u32 block_height_bit, u32 offset_x, u32 offset_y) {
-    const u32 block_height = 1U << block_height_bit;
-    const u32 image_width_in_gobs =
-        (swizzled_width * bytes_per_pixel + (GOB_SIZE_X - 1)) / GOB_SIZE_X;
-    for (u32 line = 0; line < subrect_height; ++line) {
-        const u32 dst_y = line + offset_y;
-        const u32 gob_address_y =
-            (dst_y / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height * image_width_in_gobs +
-            ((dst_y % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
-        const auto& table = SWIZZLE_TABLE[dst_y % GOB_SIZE_Y];
-        for (u32 x = 0; x < subrect_width; ++x) {
-            const u32 dst_x = x + offset_x;
-            const u32 gob_address =
-                gob_address_y + (dst_x * bytes_per_pixel / GOB_SIZE_X) * GOB_SIZE * block_height;
-            const u32 swizzled_offset = gob_address + table[(dst_x * bytes_per_pixel) % GOB_SIZE_X];
-            const u32 unswizzled_offset = line * source_pitch + x * bytes_per_pixel;
-
-            const u8* const source_line = unswizzled_data + unswizzled_offset;
-            u8* const dest_addr = swizzled_data + swizzled_offset;
-            std::memcpy(dest_addr, source_line, bytes_per_pixel);
-        }
+    switch (bytes_per_pixel) {
+#define BPP_CASE(x)                                                                                \
+    case x:                                                                                        \
+        return SwizzleSubrect<x>(subrect_width, subrect_height, source_pitch, swizzled_width,      \
+                                 swizzled_data, unswizzled_data, block_height_bit, offset_x,       \
+                                 offset_y);
+        BPP_CASE(1)
+        BPP_CASE(2)
+        BPP_CASE(3)
+        BPP_CASE(4)
+        BPP_CASE(6)
+        BPP_CASE(8)
+        BPP_CASE(12)
+        BPP_CASE(16)
+#undef BPP_CASE
+    default:
+        UNREACHABLE_MSG("Invalid bytes_per_pixel={}", bytes_per_pixel);
     }
 }
 
 void UnswizzleSubrect(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 bytes_per_pixel,
                       u32 block_height, u32 origin_x, u32 origin_y, u8* output, const u8* input) {
-    const u32 stride = width * bytes_per_pixel;
-    const u32 gobs_in_x = (stride + GOB_SIZE_X - 1) / GOB_SIZE_X;
-    const u32 block_size = gobs_in_x << (GOB_SIZE_SHIFT + block_height);
-
-    const u32 block_height_mask = (1U << block_height) - 1;
-    const u32 x_shift = GOB_SIZE_SHIFT + block_height;
-
-    for (u32 line = 0; line < line_count; ++line) {
-        const u32 src_y = line + origin_y;
-        const auto& table = SWIZZLE_TABLE[src_y % GOB_SIZE_Y];
-
-        const u32 block_y = src_y >> GOB_SIZE_Y_SHIFT;
-        const u32 src_offset_y = (block_y >> block_height) * block_size +
-                                 ((block_y & block_height_mask) << GOB_SIZE_SHIFT);
-        for (u32 column = 0; column < line_length_in; ++column) {
-            const u32 src_x = (column + origin_x) * bytes_per_pixel;
-            const u32 src_offset_x = (src_x >> GOB_SIZE_X_SHIFT) << x_shift;
-
-            const u32 swizzled_offset = src_offset_y + src_offset_x + table[src_x % GOB_SIZE_X];
-            const u32 unswizzled_offset = line * pitch + column * bytes_per_pixel;
-
-            std::memcpy(output + unswizzled_offset, input + swizzled_offset, bytes_per_pixel);
-        }
+    switch (bytes_per_pixel) {
+#define BPP_CASE(x)                                                                                \
+    case x:                                                                                        \
+        return UnswizzleSubrect<x>(line_length_in, line_count, pitch, width, block_height,         \
+                                   origin_x, origin_y, output, input);
+        BPP_CASE(1)
+        BPP_CASE(2)
+        BPP_CASE(3)
+        BPP_CASE(4)
+        BPP_CASE(6)
+        BPP_CASE(8)
+        BPP_CASE(12)
+        BPP_CASE(16)
+#undef BPP_CASE
+    default:
+        UNREACHABLE_MSG("Invalid bytes_per_pixel={}", bytes_per_pixel);
     }
 }
 
 void SwizzleSliceToVoxel(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 height,
                          u32 bytes_per_pixel, u32 block_height, u32 block_depth, u32 origin_x,
                          u32 origin_y, u8* output, const u8* input) {
-    UNIMPLEMENTED_IF(origin_x > 0);
-    UNIMPLEMENTED_IF(origin_y > 0);
-
-    const u32 stride = width * bytes_per_pixel;
-    const u32 gobs_in_x = (stride + GOB_SIZE_X - 1) / GOB_SIZE_X;
-    const u32 block_size = gobs_in_x << (GOB_SIZE_SHIFT + block_height + block_depth);
-
-    const u32 block_height_mask = (1U << block_height) - 1;
-    const u32 x_shift = static_cast<u32>(GOB_SIZE_SHIFT) + block_height + block_depth;
-
-    for (u32 line = 0; line < line_count; ++line) {
-        const auto& table = SWIZZLE_TABLE[line % GOB_SIZE_Y];
-        const u32 block_y = line / GOB_SIZE_Y;
-        const u32 dst_offset_y =
-            (block_y >> block_height) * block_size + (block_y & block_height_mask) * GOB_SIZE;
-        for (u32 x = 0; x < line_length_in; ++x) {
-            const u32 dst_offset =
-                ((x / GOB_SIZE_X) << x_shift) + dst_offset_y + table[x % GOB_SIZE_X];
-            const u32 src_offset = x * bytes_per_pixel + line * pitch;
-            std::memcpy(output + dst_offset, input + src_offset, bytes_per_pixel);
-        }
+    switch (bytes_per_pixel) {
+#define BPP_CASE(x)                                                                                \
+    case x:                                                                                        \
+        return SwizzleSliceToVoxel<x>(line_length_in, line_count, pitch, width, height,            \
+                                      block_height, block_depth, origin_x, origin_y, output,       \
+                                      input);
+        BPP_CASE(1)
+        BPP_CASE(2)
+        BPP_CASE(3)
+        BPP_CASE(4)
+        BPP_CASE(6)
+        BPP_CASE(8)
+        BPP_CASE(12)
+        BPP_CASE(16)
+#undef BPP_CASE
+    default:
+        UNREACHABLE_MSG("Invalid bytes_per_pixel={}", bytes_per_pixel);
     }
 }
 
@@ -228,7 +287,7 @@ void SwizzleKepler(const u32 width, const u32 height, const u32 dst_x, const u32
             u8* dest_addr = swizzle_data + swizzled_offset;
             count++;
 
-            std::memcpy(dest_addr, source_line, 1);
+            *dest_addr = *source_line;
         }
     }
 }