[Mesa-dev] [PATCH 10/13] i965/miptree: Use cpu tiling/detiling when mapping

Scott D Phillips scott.d.phillips at intel.com
Mon Apr 30 23:38:57 UTC 2018


Kenneth Graunke <kenneth at whitecape.org> writes:

> On Monday, April 30, 2018 10:25:49 AM PDT Scott D Phillips wrote:
>> Rename the (un)map_gtt functions to (un)map_map (map by
>> returning a map) and add new functions (un)map_tiled_memcpy that
>> return a shadow buffer populated with the intel_tiled_memcpy
>> functions.
>> 
>> Tiling/detiling with the cpu will be the only way to handle Yf/Ys
>> tiling, when support is added for those formats.
>> 
>> v2: Compute extents properly in the x|y-rounded-down case (Chris Wilson)
>> 
>> v3: Add units to parameter names of tile_extents (Nanley Chery)
>>     Use _mesa_align_malloc for the shadow copy (Nanley)
>>     Continue using gtt maps on gen4 (Nanley)
>> 
>> v4: Use streaming_load_memcpy when detiling
>> 
>> Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>
>> ---
>>  src/mesa/drivers/dri/i965/intel_mipmap_tree.c | 98 +++++++++++++++++++++++++--
>>  1 file changed, 94 insertions(+), 4 deletions(-)
>> 
>> diff --git a/src/mesa/drivers/dri/i965/intel_mipmap_tree.c b/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
>> index b9a564552df..498eebd2f86 100644
>> --- a/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
>> +++ b/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
>> @@ -31,6 +31,7 @@
>>  #include "intel_image.h"
>>  #include "intel_mipmap_tree.h"
>>  #include "intel_tex.h"
>> +#include "intel_tiled_memcpy.h"
>>  #include "intel_blit.h"
>>  #include "intel_fbo.h"
>>  
>> @@ -3066,7 +3067,7 @@ intel_miptree_unmap_raw(struct intel_mipmap_tree *mt)
>>  }
>>  
>>  static void
>> -intel_miptree_unmap_gtt(struct brw_context *brw,
>> +intel_miptree_unmap_map(struct brw_context *brw,
>>                          struct intel_mipmap_tree *mt,
>>                          struct intel_miptree_map *map,
>>                          unsigned int level, unsigned int slice)
>> @@ -3075,7 +3076,7 @@ intel_miptree_unmap_gtt(struct brw_context *brw,
>>  }
>>  
>>  static void
>> -intel_miptree_map_gtt(struct brw_context *brw,
>> +intel_miptree_map_map(struct brw_context *brw,
>>  		      struct intel_mipmap_tree *mt,
>>  		      struct intel_miptree_map *map,
>>  		      unsigned int level, unsigned int slice)
>> @@ -3120,7 +3121,7 @@ intel_miptree_map_gtt(struct brw_context *brw,
>>         mt, _mesa_get_format_name(mt->format),
>>         x, y, map->ptr, map->stride);
>>  
>> -   map->unmap = intel_miptree_unmap_gtt;
>> +   map->unmap = intel_miptree_unmap_map;
>>  }
>>  
>>  static void
>> @@ -3145,6 +3146,90 @@ intel_miptree_unmap_blit(struct brw_context *brw,
>>     intel_miptree_release(&map->linear_mt);
>>  }
>>  
>> +/* Compute extent parameters for use with tiled_memcpy functions.
>> + * xs are in units of bytes and ys are in units of strides. */
>> +static inline void
>> +tile_extents(struct intel_mipmap_tree *mt, struct intel_miptree_map *map,
>> +             unsigned int level, unsigned int slice, unsigned int *x1_B,
>> +             unsigned int *x2_B, unsigned int *y1_el, unsigned int *y2_el)
>> +{
>> +   unsigned int block_width, block_height;
>> +   unsigned int x0_el, y0_el;
>> +
>> +   _mesa_get_format_block_size(mt->format, &block_width, &block_height);
>> +
>> +   assert(map->x % block_width == 0);
>> +   assert(map->y % block_height == 0);
>> +
>> +   intel_miptree_get_image_offset(mt, level, slice, &x0_el, &y0_el);
>> +   *x1_B = (map->x / block_width + x0_el) * mt->cpp;
>> +   *y1_el = map->y / block_height + y0_el;
>> +   *x2_B = (DIV_ROUND_UP(map->x + map->w, block_width) + x0_el) * mt->cpp;
>> +   *y2_el = DIV_ROUND_UP(map->y + map->h, block_height) + y0_el;
>> +}
>> +
>> +static void
>> +intel_miptree_unmap_tiled_memcpy(struct brw_context *brw,
>> +                                 struct intel_mipmap_tree *mt,
>> +                                 struct intel_miptree_map *map,
>> +                                 unsigned int level,
>> +                                 unsigned int slice)
>> +{
>> +   if (map->mode & GL_MAP_WRITE_BIT) {
>> +      unsigned int x1, x2, y1, y2;
>> +      tile_extents(mt, map, level, slice, &x1, &x2, &y1, &y2);
>> +
>> +      char *dst = intel_miptree_map_raw(brw, mt, map->mode | MAP_RAW);
>> +      dst += mt->offset;
>> +
>> +      linear_to_tiled(x1, x2, y1, y2, dst, map->ptr, mt->surf.row_pitch,
>> +                      map->stride, brw->has_swizzling, mt->surf.tiling, memcpy);
>> +
>> +      intel_miptree_unmap_raw(mt);
>> +   }
>> +   _mesa_align_free(map->buffer);
>> +   map->buffer = map->ptr = NULL;
>> +}
>> +
>> +static void
>> +intel_miptree_map_tiled_memcpy(struct brw_context *brw,
>> +                               struct intel_mipmap_tree *mt,
>> +                               struct intel_miptree_map *map,
>> +                               unsigned int level, unsigned int slice)
>> +{
>> +   unsigned int x1, x2, y1, y2;
>> +   tile_extents(mt, map, level, slice, &x1, &x2, &y1, &y2);
>> +   map->stride = ALIGN(_mesa_format_row_stride(mt->format, map->w), 16);
>> +
>> +   /* The tiling and detiling functions require that the linear buffer
>> +    * has proper 16-byte alignment (that is, its `x0` is 16-byte
>> +    * aligned). Here we over-allocate the linear buffer by enough
>> +    * bytes to get the proper alignment.
>> +    */
>> +   map->buffer = _mesa_align_malloc(map->stride * (y2 - y1) + (x1 & 0xf), 16);
>> +   map->ptr = (char *)map->buffer + (x1 & 0xf);
>> +   assert(map->buffer);
>> +
>> +   if (!(map->mode & GL_MAP_INVALIDATE_RANGE_BIT)) {
>> +      char *src = intel_miptree_map_raw(brw, mt, map->mode | MAP_RAW);
>> +      src += mt->offset;
>> +
>> +      const mem_copy_fn fn =
>> +#if defined(USE_SSE41)
>> +         cpu_has_sse4_1 ? (mem_copy_fn)_mesa_streaming_load_memcpy :
>> +#endif
>> +         memcpy;
>> +
>> +      tiled_to_linear(x1, x2, y1, y2, map->ptr, src, map->stride,
>> +                      mt->surf.row_pitch, brw->has_swizzling, mt->surf.tiling,
>> +                      fn);
>> +
>> +      intel_miptree_unmap_raw(mt);
>> +   }
>> +
>> +   map->unmap = intel_miptree_unmap_tiled_memcpy;
>> +}
>> +
>>  static void
>>  intel_miptree_map_blit(struct brw_context *brw,
>>  		       struct intel_mipmap_tree *mt,
>> @@ -3688,8 +3773,13 @@ intel_miptree_map(struct brw_context *brw,
>>                (mt->surf.row_pitch % 16 == 0)) {
>>        intel_miptree_map_movntdqa(brw, mt, map, level, slice);
>>  #endif
>> +   } else if (mt->surf.tiling != ISL_TILING_LINEAR &&
>> +              brw->screen->devinfo.gen > 4) {
>> +      intel_miptree_map_tiled_memcpy(brw, mt, map, level, slice);
>>     } else {
>> -      intel_miptree_map_gtt(brw, mt, map, level, slice);
>> +      if (mt->surf.tiling != ISL_TILING_LINEAR)
>> +         perf_debug("intel_miptree_map: mapping via gtt");
>> +      intel_miptree_map_map(brw, mt, map, level, slice);
>>     }
>>  
>>     *out_ptr = map->ptr;
>> 
>
> I'm not exactly sure what BRW_MAP_DIRECT_BIT is supposed to mean, or how
> useful it is.  It seems like it means "directly map the underlying
> surface, without making temporaries".  If so, we should probably skip
> this path if it's set.
>
> But, in the last patch, you set BRW_MAP_DIRECT_BIT to avoid the
> intel_miptree_map_depthstencil case when recursing...yet you want to
> use the tiled_memcpy paths.
>
> So, I'm not sure what to suggest.

I guess it's a bit messy. I was taking MAP_DIRECT_BIT to mean "don't
reformat the surface," and then we also have MAP_RAW which I would take
to mean "just give me the memory." So maybe I should be checking for
MAP_RAW in this tree and skipping the tiling in that case? Or maybe
rename DIRECT? Not sure.

> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list