<div dir="ltr">Reviewed-by: Jason Ekstrand <<a href="mailto:jason@jlekstrand.net">jason@jlekstrand.net</a>> </div><div class="gmail_extra"> <div class="gmail_quote">On Mon, Apr 30, 2018 at 10:25 AM, Scott D Phillips <<a href="mailto:scott.d.phillips@intel.com" target="_blank">scott.d.phillips@intel.com</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Similar to the transformation applied to linear_to_ytiled, also align each readback from the ytiled source to a cacheline (i.e. transfer a whole cacheline from the source before moving on to the next column). This will allow us to utilize movntqda (_mm_stream_si128) in a subsequent patch to obtain near WB readback performance when accessing the uncached ytiled memory, an order of magnitude improvement. Reviewed-by: Chris Wilson <<a href="mailto:chris@chris-wilson.co.uk">chris@chris-wilson.co.uk</a>> --- src/mesa/drivers/dri/i965/intel_tiled_memcpy.c | 72 +++++++++++++++++++++++--- 1 file changed, 66 insertions(+), 6 deletions(-) diff --git a/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c b/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c index 69306828d72..7c6bde990d6 100644 --- a/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c +++ b/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c @@ -451,7 +451,7 @@ xtiled_to_linear(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3, */ static inline void ytiled_to_linear(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3, - uint32_t y0, uint32_t y1, + uint32_t y0, uint32_t y3, char *dst, const char *src, int32_t dst_pitch, uint32_t swizzle_bit, @@ -470,6 +470,9 @@ ytiled_to_linear(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3, const uint32_t column_width = ytile_span; const uint32_t bytes_per_column = column_width * ytile_height; + uint32_t y1 = MIN2(y3, ALIGN_UP(y0, 4)); + uint32_t y2 = MAX2(y1, ALIGN_DOWN(y3, 4)); + uint32_t xo0 = (x0 % ytile_span) + (x0 / ytile_span) * bytes_per_column; uint32_t xo1 = (x1 % ytile_span) + (x1 / ytile_span) * bytes_per_column; @@ -485,24 +488,81 @@ ytiled_to_linear(uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3, dst += (ptrdiff_t)y0 * dst_pitch; - for (yo = y0 * column_width; yo < y1 * column_width; yo += column_width) { + if (y0 != y1) { + for (yo = y0 * column_width; yo < y1 * column_width; yo += column_width) { + uint32_t xo = xo1; + uint32_t swizzle = swizzle1; + + mem_copy(dst + x0, src + ((xo0 + yo) ^ swizzle0), 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_align16(dst + x, src + ((xo + yo) ^ swizzle), ytile_span); + xo += bytes_per_column; + swizzle ^= swizzle_bit; + } + + mem_copy_align16(dst + x2, src + ((xo + yo) ^ swizzle), x3 - x2); + + dst += dst_pitch; + } + } + + for (yo = y1 * column_width; yo < y2 * column_width; yo += 4 * column_width) { uint32_t xo = xo1; uint32_t swizzle = swizzle1; - mem_copy(dst + x0, src + ((xo0 + yo) ^ swizzle0), x1 - x0); + if (x0 != x1) { + mem_copy(dst + x0 + 0 * dst_pitch, src + ((xo0 + yo + 0 * column_width) ^ swizzle0), x1 - x0); + mem_copy(dst + x0 + 1 * dst_pitch, src + ((xo0 + yo + 1 * column_width) ^ swizzle0), x1 - x0); + mem_copy(dst + x0 + 2 * dst_pitch, src + ((xo0 + yo + 2 * column_width) ^ swizzle0), x1 - x0); + mem_copy(dst + x0 + 3 * dst_pitch, src + ((xo0 + yo + 3 * column_width) ^ swizzle0), 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_align16(dst + x, src + ((xo + yo) ^ swizzle), ytile_span); + mem_copy_align16(dst + x + 0 * dst_pitch, src + ((xo + yo + 0 * column_width) ^ swizzle), ytile_span); + mem_copy_align16(dst + x + 1 * dst_pitch, src + ((xo + yo + 1 * column_width) ^ swizzle), ytile_span); + mem_copy_align16(dst + x + 2 * dst_pitch, src + ((xo + yo + 2 * column_width) ^ swizzle), ytile_span); + mem_copy_align16(dst + x + 3 * dst_pitch, src + ((xo + yo + 3 * column_width) ^ swizzle), ytile_span); xo += bytes_per_column; swizzle ^= swizzle_bit; } - mem_copy_align16(dst + x2, src + ((xo + yo) ^ swizzle), x3 - x2); + if (x2 != x3) { + mem_copy_align16(dst + x2 + 0 * dst_pitch, src + ((xo + yo + 0 * column_width) ^ swizzle), x3 - x2); + mem_copy_align16(dst + x2 + 1 * dst_pitch, src + ((xo + yo + 1 * column_width) ^ swizzle), x3 - x2); + mem_copy_align16(dst + x2 + 2 * dst_pitch, src + ((xo + yo + 2 * column_width) ^ swizzle), x3 - x2); + mem_copy_align16(dst + x2 + 3 * dst_pitch, src + ((xo + yo + 3 * column_width) ^ swizzle), x3 - x2); + } - dst += dst_pitch; + dst += 4 * dst_pitch; + } + + if (y2 != y3) { + for (yo = y2 * column_width; yo < y3 * column_width; yo += column_width) { + uint32_t xo = xo1; + uint32_t swizzle = swizzle1; + + mem_copy(dst + x0, src + ((xo0 + yo) ^ swizzle0), 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_align16(dst + x, src + ((xo + yo) ^ swizzle), ytile_span); + xo += bytes_per_column; + swizzle ^= swizzle_bit; + } + + mem_copy_align16(dst + x2, src + ((xo + yo) ^ swizzle), x3 - x2); + + dst += dst_pitch; + } } } -- 2.14.3 </blockquote></div> </div>