1 files changed, 402 insertions, 69 deletions

diff --git a/src/mesa/drivers/dri/i965/intel_tex_subimage.c b/src/mesa/drivers/dri/i965/intel_tex_subimage.c
index 05e684cf57b..5cfdbd96824 100644
--- a/src/mesa/drivers/dri/i965/intel_tex_subimage.c
+++ b/src/mesa/drivers/dri/i965/intel_tex_subimage.c
@@ -43,6 +43,43 @@
 
 #define FILE_DEBUG_FLAG DEBUG_TEXTURE
 
+#define ALIGN_DOWN(a, b) ROUND_DOWN_TO(a, b)
+#define ALIGN_UP(a, b) ALIGN(a, b)
+
+/* Tile dimensions.
+ * Width and span are in bytes, height is in pixels (i.e. unitless).
+ * A "span" is the most number of bytes we can copy from linear to tiled
+ * without needing to calculate a new destination address.
+ */
+static const uint32_t xtile_width = 512;
+static const uint32_t xtile_height = 8;
+static const uint32_t xtile_span = 64;
+static const uint32_t ytile_width = 128;
+static const uint32_t ytile_height = 32;
+static const uint32_t ytile_span = 16;
+
+typedef void *(*mem_copy_fn)(void *dest, const void *src, size_t n);
+
+/**
+ * Each row from y0 to y1 is copied in three parts: [x0,x1), [x1,x2), [x2,x3).
+ * These ranges are in bytes, i.e. pixels * bytes-per-pixel.
+ * The first and last ranges must be shorter than a "span" (the longest linear
+ * stretch within a tile) and the middle must equal a whole number of spans.
+ * Ranges may be empty.  The region copied must land entirely within one tile.
+ * 'dst' is the start of the tile and 'src' is the corresponding
+ * address to copy from, though copying begins at (x0, y0).
+ * To enable swizzling 'swizzle_bit' must be 1<<6, otherwise zero.
+ * Swizzling flips bit 6 in the copy destination offset, when certain other
+ * bits are set in it.
+ */
+typedef void (*tile_copy_fn)(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
+                             uint32_t y0, uint32_t y1,
+                             char *dst, const char *src,
+                             uint32_t src_pitch,
+                             uint32_t swizzle_bit,
+                             mem_copy_fn mem_copy);
+
+
 static bool
 intel_blit_texsubimage(struct gl_context * ctx,
 		       struct gl_texture_image *texImage,
@@ -132,26 +169,351 @@ err:
    return false;
 }
 
+#ifdef __SSSE3__
+static const uint8_t rgba8_permutation[16] =
+   { 2,1,0,3, 6,5,4,7, 10,9,8,11, 14,13,12,15 };
+
+typedef char v16 __attribute__((vector_size(16)));
+
+/* NOTE: dst must be 16 byte aligned */
+#define rgba8_copy_16(dst, src)                     \
+   *(v16*)(dst) = __builtin_ia32_pshufb128(         \
+       (v16) __builtin_ia32_loadups((float*)(src)), \
+      *(v16*) rgba8_permutation                     \
+   )
+#endif
+
+/**
+ * Copy RGBA to BGRA - swap R and B.
+ */
+static inline void *
+rgba8_copy(void *dst, const void *src, size_t bytes)
+{
+   uint8_t *d = dst;
+   uint8_t const *s = src;
+
+#ifdef __SSSE3__
+   /* Fast copying for tile spans.
+    *
+    * As long as the destination texture is 16 aligned,
+    * any 16 or 64 spans we get here should also be 16 aligned.
+    */
+
+   if (bytes == 16) {
+      assert(!(((uintptr_t)dst) & 0xf));
+      rgba8_copy_16(d+ 0, s+ 0);
+      return dst;
+   }
+
+   if (bytes == 64) {
+      assert(!(((uintptr_t)dst) & 0xf));
+      rgba8_copy_16(d+ 0, s+ 0);
+      rgba8_copy_16(d+16, s+16);
+      rgba8_copy_16(d+32, s+32);
+      rgba8_copy_16(d+48, s+48);
+      return dst;
+   }
+#endif
+
+   while (bytes >= 4) {
+      d[0] = s[2];
+      d[1] = s[1];
+      d[2] = s[0];
+      d[3] = s[3];
+      d += 4;
+      s += 4;
+      bytes -= 4;
+   }
+   return dst;
+}
+
+/**
+ * Copy texture data from linear to X tile layout.
+ *
+ * \copydoc tile_copy_fn
+ */
+static inline void
+xtile_copy(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
+           uint32_t y0, uint32_t y1,
+           char *dst, const char *src,
+           uint32_t src_pitch,
+           uint32_t swizzle_bit,
+           mem_copy_fn mem_copy)
+{
+   /* The copy destination offset for each range copied is the sum of
+    * an X offset 'x0' or 'xo' and a Y offset 'yo.'
+    */
+   uint32_t xo, yo;
+
+   src += y0 * src_pitch;
+
+   for (yo = y0 * xtile_width; yo < y1 * xtile_width; yo += xtile_width) {
+      /* Bits 9 and 10 of the copy destination offset control swizzling.
+       * Only 'yo' contributes to those bits in the total offset,
+       * so calculate 'swizzle' just once per row.
+       * Move bits 9 and 10 three and four places respectively down
+       * to bit 6 and xor them.
+       */
+      uint32_t swizzle = ((yo >> 3) ^ (yo >> 4)) & swizzle_bit;
+
+      mem_copy(dst + ((x0 + yo) ^ swizzle), src + x0, x1 - x0);
+
+      for (xo = x1; xo < x2; xo += xtile_span) {
+         mem_copy(dst + ((xo + yo) ^ swizzle), src + xo, xtile_span);
+      }
+
+      mem_copy(dst + ((xo + yo) ^ swizzle), src + x2, x3 - x2);
+
+      src += src_pitch;
+   }
+}
+
+/**
+ * Copy texture data from linear to Y tile layout.
+ *
+ * \copydoc tile_copy_fn
+ */
+static inline void
+ytile_copy(
+   uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
+   uint32_t y0, uint32_t y1,
+   char *dst, const char *src,
+   uint32_t src_pitch,
+   uint32_t swizzle_bit,
+   mem_copy_fn mem_copy)
+{
+   /* Y tiles consist of columns that are 'ytile_span' wide (and the same height
+    * as the tile).  Thus the destination offset for (x,y) is the sum of:
+    *   (x % column_width)                    // position within column
+    *   (x / column_width) * bytes_per_column // column number * bytes per column
+    *   y * column_width
+    *
+    * The copy destination offset for each range copied is the sum of
+    * an X offset 'xo0' or 'xo' and a Y offset 'yo.'
+    */
+   const uint32_t column_width = ytile_span;
+   const uint32_t bytes_per_column = column_width * ytile_height;
+
+   uint32_t xo0 = (x0 % ytile_span) + (x0 / ytile_span) * bytes_per_column;
+   uint32_t xo1 = (x1 % ytile_span) + (x1 / ytile_span) * bytes_per_column;
+
+   /* Bit 9 of the destination offset control swizzling.
+    * Only the X offset contributes to bit 9 of the total offset,
+    * so swizzle can be calculated in advance for these X positions.
+    * Move bit 9 three places down to bit 6.
+    */
+   uint32_t swizzle0 = (xo0 >> 3) & swizzle_bit;
+   uint32_t swizzle1 = (xo1 >> 3) & swizzle_bit;
+
+   uint32_t x, yo;
+
+   src += y0 * src_pitch;
+
+   for (yo = y0 * column_width; yo < y1 * column_width; yo += column_width) {
+      uint32_t xo = xo1;
+      uint32_t swizzle = swizzle1;
+
+      mem_copy(dst + ((xo0 + yo) ^ swizzle0), src + x0, x1 - x0);
+
+      /* Step by spans/columns.  As it happens, the swizzle bit flips
+       * at each step so we don't need to calculate it explicitly.
+       */
+      for (x = x1; x < x2; x += ytile_span) {
+         mem_copy(dst + ((xo + yo) ^ swizzle), src + x, ytile_span);
+         xo += bytes_per_column;
+         swizzle ^= swizzle_bit;
+      }
+
+      mem_copy(dst + ((xo + yo) ^ swizzle), src + x2, x3 - x2);
+
+      src += src_pitch;
+   }
+}
+
+/**
+ * Copy texture data from linear to X tile layout, faster.
+ *
+ * Same as \ref xtile_copy but faster, because it passes constant parameters
+ * for common cases, allowing the compiler to inline code optimized for those
+ * cases.
+ *
+ * \copydoc tile_copy_fn
+ */
+static void
+xtile_copy_faster(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
+                  uint32_t y0, uint32_t y1,
+                  char *dst, const char *src,
+                  uint32_t src_pitch,
+                  uint32_t swizzle_bit,
+                  mem_copy_fn mem_copy)
+{
+   if (x0 == 0 && x3 == xtile_width && y0 == 0 && y1 == xtile_height) {
+      if (mem_copy == memcpy)
+         return xtile_copy(0, 0, xtile_width, xtile_width, 0, xtile_height,
+                           dst, src, src_pitch, swizzle_bit, memcpy);
+      else if (mem_copy == rgba8_copy)
+         return xtile_copy(0, 0, xtile_width, xtile_width, 0, xtile_height,
+                           dst, src, src_pitch, swizzle_bit, rgba8_copy);
+   } else {
+      if (mem_copy == memcpy)
+         return xtile_copy(x0, x1, x2, x3, y0, y1,
+                           dst, src, src_pitch, swizzle_bit, memcpy);
+      else if (mem_copy == rgba8_copy)
+         return xtile_copy(x0, x1, x2, x3, y0, y1,
+                           dst, src, src_pitch, swizzle_bit, rgba8_copy);
+   }
+   xtile_copy(x0, x1, x2, x3, y0, y1,
+              dst, src, src_pitch, swizzle_bit, mem_copy);
+}
+
+/**
+ * Copy texture data from linear to Y tile layout, faster.
+ *
+ * Same as \ref ytile_copy but faster, because it passes constant parameters
+ * for common cases, allowing the compiler to inline code optimized for those
+ * cases.
+ *
+ * \copydoc tile_copy_fn
+ */
+static void
+ytile_copy_faster(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3,
+                  uint32_t y0, uint32_t y1,
+                  char *dst, const char *src,
+                  uint32_t src_pitch,
+                  uint32_t swizzle_bit,
+                  mem_copy_fn mem_copy)
+{
+   if (x0 == 0 && x3 == ytile_width && y0 == 0 && y1 == ytile_height) {
+      if (mem_copy == memcpy)
+         return ytile_copy(0, 0, ytile_width, ytile_width, 0, ytile_height,
+                           dst, src, src_pitch, swizzle_bit, memcpy);
+      else if (mem_copy == rgba8_copy)
+         return ytile_copy(0, 0, ytile_width, ytile_width, 0, ytile_height,
+                           dst, src, src_pitch, swizzle_bit, rgba8_copy);
+   } else {
+      if (mem_copy == memcpy)
+         return ytile_copy(x0, x1, x2, x3, y0, y1,
+                           dst, src, src_pitch, swizzle_bit, memcpy);
+      else if (mem_copy == rgba8_copy)
+         return ytile_copy(x0, x1, x2, x3, y0, y1,
+                           dst, src, src_pitch, swizzle_bit, rgba8_copy);
+   }
+   ytile_copy(x0, x1, x2, x3, y0, y1,
+              dst, src, src_pitch, swizzle_bit, mem_copy);
+}
+
+/**
+ * Copy from linear to tiled texture.
+ *
+ * Divide the region given by X range [xt1, xt2) and Y range [yt1, yt2) into
+ * pieces that do not cross tile boundaries and copy each piece with a tile
+ * copy function (\ref tile_copy_fn).
+ * The X range is in bytes, i.e. pixels * bytes-per-pixel.
+ * The Y range is in pixels (i.e. unitless).
+ * 'dst' is the start of the texture and 'src' is the corresponding
+ * address to copy from, though copying begins at (xt1, yt1).
+ */
+static void
+linear_to_tiled(uint32_t xt1, uint32_t xt2,
+                uint32_t yt1, uint32_t yt2,
+                char *dst, const char *src,
+                uint32_t dst_pitch, uint32_t src_pitch,
+                bool has_swizzling,
+                uint32_t tiling,
+                mem_copy_fn mem_copy)
+{
+   tile_copy_fn tile_copy;
+   uint32_t xt0, xt3;
+   uint32_t yt0, yt3;
+   uint32_t xt, yt;
+   uint32_t tw, th, span;
+   uint32_t swizzle_bit = has_swizzling ? 1<<6 : 0;
+
+   if (tiling == I915_TILING_X) {
+      tw = xtile_width;
+      th = xtile_height;
+      span = xtile_span;
+      tile_copy = xtile_copy_faster;
+   } else if (tiling == I915_TILING_Y) {
+      tw = ytile_width;
+      th = ytile_height;
+      span = ytile_span;
+      tile_copy = ytile_copy_faster;
+   } else {
+      assert(!"unsupported tiling");
+      return;
+   }
+
+   /* Round out to tile boundaries. */
+   xt0 = ALIGN_DOWN(xt1, tw);
+   xt3 = ALIGN_UP  (xt2, tw);
+   yt0 = ALIGN_DOWN(yt1, th);
+   yt3 = ALIGN_UP  (yt2, th);
+
+   /* Loop over all tiles to which we have something to copy.
+    * 'xt' and 'yt' are the origin of the destination tile, whether copying
+    * copying a full or partial tile.
+    * tile_copy() copies one tile or partial tile.
+    * Looping x inside y is the faster memory access pattern.
+    */
+   for (yt = yt0; yt < yt3; yt += th) {
+      for (xt = xt0; xt < xt3; xt += tw) {
+         /* The area to update is [x0,x3) x [y0,y1).
+          * May not want the whole tile, hence the min and max.
+          */
+         uint32_t x0 = MAX2(xt1, xt);
+         uint32_t y0 = MAX2(yt1, yt);
+         uint32_t x3 = MIN2(xt2, xt + tw);
+         uint32_t y1 = MIN2(yt2, yt + th);
+
+         /* [x0,x3) is split into [x0,x1), [x1,x2), [x2,x3) such that
+          * the middle interval is the longest span-aligned part.
+          * The sub-ranges could be empty.
+          */
+         uint32_t x1, x2;
+         x1 = ALIGN_UP(x0, span);
+         if (x1 > x3)
+            x1 = x2 = x3;
+         else
+            x2 = ALIGN_DOWN(x3, span);
+
+         assert(x0 <= x1 && x1 <= x2 && x2 <= x3);
+         assert(x1 - x0 < span && x3 - x2 < span);
+         assert(x3 - x0 <= tw);
+         assert((x2 - x1) % span == 0);
+
+         /* Translate by (xt,yt) for single-tile copier. */
+         tile_copy(x0-xt, x1-xt, x2-xt, x3-xt,
+                   y0-yt, y1-yt,
+                   dst + xt * th + yt * dst_pitch,
+                   src + xt      + yt * src_pitch,
+                   src_pitch,
+                   swizzle_bit,
+                   mem_copy);
+      }
+   }
+}
+
 /**
  * \brief A fast path for glTexImage and glTexSubImage.
  *
  * \param for_glTexImage Was this called from glTexImage or glTexSubImage?
  *
- * This fast path is taken when the hardware natively supports the texture
- * format (such as GL_BGRA) and when the texture memory is X-tiled. It uploads
+ * This fast path is taken when the texture format is BGRA, RGBA,
+ * A or L and when the texture memory is X- or Y-tiled.  It uploads
  * the texture data by mapping the texture memory without a GTT fence, thus
- * acquiring a tiled view of the memory, and then memcpy'ing sucessive
- * subspans within each tile.
+ * acquiring a tiled view of the memory, and then copying sucessive
+ * spans within each tile.
  *
  * This is a performance win over the conventional texture upload path because
  * it avoids the performance penalty of writing through the write-combine
  * buffer. In the conventional texture upload path,
  * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
  * fence, thus acquiring a linear view of the memory, then each row in the
- * image is memcpy'd. In this fast path, we replace each row's memcpy with
- * a sequence of memcpy's over each bit6 swizzle span in the row.
+ * image is memcpy'd. In this fast path, we replace each row's copy with
+ * a sequence of copies over each linear span in tile.
  *
- * This fast path's use case is Google Chrome's paint rectangles.  Chrome (as
+ * One use case is Google Chrome's paint rectangles.  Chrome (as
  * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
  * Each page's content is initially uploaded with glTexImage2D and damaged
  * regions are updated with glTexSubImage2D. On some workloads, the
@@ -176,14 +538,15 @@ intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
 
    int error = 0;
 
-   /* This fastpath is restricted to a specific texture type: level 0 of
-    * a 2D BGRA texture. It could be generalized to support more types by
-    * varying the arithmetic loop below.
+   uint32_t cpp;
+   mem_copy_fn mem_copy = NULL;
+
+   /* This fastpath is restricted to specific texture types: level 0 of
+    * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
+    * more types.
     */
    if (!brw->has_llc ||
-       format != GL_BGRA ||
        type != GL_UNSIGNED_BYTE ||
-       texImage->TexFormat != MESA_FORMAT_ARGB8888 ||
        texImage->TexObject->Target != GL_TEXTURE_2D ||
        texImage->Level != 0 ||
        pixels == NULL ||
@@ -197,12 +560,28 @@ intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
        packing->Invert)
       return false;
 
+   if ((texImage->TexFormat == MESA_FORMAT_L8 && format == GL_LUMINANCE) ||
+       (texImage->TexFormat == MESA_FORMAT_A8 && format == GL_ALPHA)) {
+      cpp = 1;
+      mem_copy = memcpy;
+   } else if (texImage->TexFormat == MESA_FORMAT_ARGB8888) {
+      cpp = 4;
+      if (format == GL_BGRA) {
+         mem_copy = memcpy;
+      } else if (format == GL_RGBA) {
+         mem_copy = rgba8_copy;
+      }
+   }
+   if (!mem_copy)
+      return false;
+
    if (for_glTexImage)
       ctx->Driver.AllocTextureImageBuffer(ctx, texImage);
 
    if (!image->mt ||
-       image->mt->region->tiling != I915_TILING_X) {
-      /* The algorithm below is written only for X-tiled memory. */
+       (image->mt->region->tiling != I915_TILING_X &&
+       image->mt->region->tiling != I915_TILING_Y)) {
+      /* The algorithm is written only for X- or Y-tiled memory. */
       return false;
    }
 
@@ -236,61 +615,15 @@ intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
    DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d)\n",
        __FUNCTION__, texImage->Level, xoffset, yoffset, width, height);
 
-   /* In the tiling algorithm below, some variables are in units of pixels,
-    * others are in units of bytes, and others (such as height) are unitless.
-    * Each variable name is suffixed with its units.
-    */
-
-   const uint32_t x_max_pixels = xoffset + width;
-   const uint32_t y_max_pixels = yoffset + height;
-
-   const uint32_t tile_size_bytes = 4096;
-
-   const uint32_t tile_width_bytes = 512;
-   const uint32_t tile_width_pixels = 128;
-
-   const uint32_t tile_height = 8;
-
-   const uint32_t cpp = 4; /* chars per pixel of GL_BGRA */
-   const uint32_t swizzle_width_pixels = 16;
-
-   const uint32_t stride_bytes = image->mt->region->pitch;
-   const uint32_t width_tiles = stride_bytes / tile_width_bytes;
-
-   for (uint32_t y_pixels = yoffset; y_pixels < y_max_pixels; ++y_pixels) {
-      const uint32_t y_offset_bytes = (y_pixels / tile_height) * width_tiles * tile_size_bytes
-                                    + (y_pixels % tile_height) * tile_width_bytes;
-
-      for (uint32_t x_pixels = xoffset; x_pixels < x_max_pixels; x_pixels += swizzle_width_pixels) {
-         const uint32_t x_offset_bytes = (x_pixels / tile_width_pixels) * tile_size_bytes
-                                       + (x_pixels % tile_width_pixels) * cpp;
-
-         intptr_t offset_bytes = y_offset_bytes + x_offset_bytes;
-         if (brw->has_swizzling) {
-#if 0
-            /* Clear, unoptimized version. */
-            bool bit6 = (offset_bytes >> 6) & 1;
-            bool bit9 = (offset_bytes >> 9) & 1;
-            bool bit10 = (offset_bytes >> 10) & 1;
-
-            if (bit9 ^ bit10)
-               offset_bytes ^= (1 << 6);
-#else
-            /* Optimized, obfuscated version. */
-            offset_bytes ^= ((offset_bytes >> 3) ^ (offset_bytes >> 4))
-                          & (1 << 6);
-#endif
-         }
-
-         const uint32_t swizzle_bound_pixels = ALIGN(x_pixels + 1, swizzle_width_pixels);
-         const uint32_t memcpy_bound_pixels = MIN2(x_max_pixels, swizzle_bound_pixels);
-         const uint32_t copy_size = cpp * (memcpy_bound_pixels - x_pixels);
-
-         memcpy(bo->virtual + offset_bytes, pixels, copy_size);
-         pixels += copy_size;
-         x_pixels -= (x_pixels % swizzle_width_pixels);
-      }
-   }
+   linear_to_tiled(
+      xoffset * cpp, (xoffset + width) * cpp,
+      yoffset, yoffset + height,
+      bo->virtual, pixels - (xoffset + yoffset * width) * cpp,
+      image->mt->region->pitch, width * cpp,
+      brw->has_swizzling,
+      image->mt->region->tiling,
+      mem_copy
+   );
 
    drm_intel_bo_unmap(bo);
    return true;