From: Jonathan Marek Date: Tue, 9 Jun 2020 22:48:34 +0000 (-0400) Subject: freedreno/fdl6: rework layout code a bit (reduce linear align to 64 bytes) X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=d53dc6c37680eba8e8fedde055e03b0a61407467;p=mesa.git freedreno/fdl6: rework layout code a bit (reduce linear align to 64 bytes) Reduce linear alignment, and rework the layout code a bit. This rework has a side effect of also increasing the alignment on linear levels of tiled (non-ubwc) cpp=1 and cpp=2 layouts. Since we should be UBWC for those cases anyway, its not a big loss. Signed-off-by: Jonathan Marek Part-of: --- diff --git a/src/freedreno/fdl/fd6_layout.c b/src/freedreno/fdl/fd6_layout.c index 146e86900aa..b5ffde5f8fe 100644 --- a/src/freedreno/fdl/fd6_layout.c +++ b/src/freedreno/fdl/fd6_layout.c @@ -29,62 +29,70 @@ #include "freedreno_layout.h" -/* indexed by cpp, including msaa 2x and 4x: - * TODO: - * cpp=1 UBWC needs testing at larger texture sizes - * missing UBWC blockwidth/blockheight for npot+64 cpp - * missing 96/128 CPP for 8x MSAA with 32_32_32/32_32_32_32 - */ -static const struct tile_alignment { - unsigned basealign; - unsigned pitchalign; - unsigned heightalign; - /* UBWC block width/height. Used in size alignment, and calculating a - * descriptor's FLAG_BUFFER_LOG2W/H for mipmapping. - */ - uint8_t ubwc_blockwidth; - uint8_t ubwc_blockheight; -} tile_alignment[] = { - [1] = { 64, 128, 32, 16, 4 }, - [2] = { 128, 128, 16, 16, 4 }, - [3] = { 256, 64, 32 }, - [4] = { 256, 64, 16, 16, 4 }, - [6] = { 256, 64, 16 }, - [8] = { 256, 64, 16, 8, 4, }, - [12] = { 256, 64, 16 }, - [16] = { 256, 64, 16, 4, 4, }, - [24] = { 256, 64, 16 }, - [32] = { 256, 64, 16, 4, 2 }, - [48] = { 256, 64, 16 }, - [64] = { 256, 64, 16 }, - - /* special cases for r8g8: */ - [0] = { 256, 64, 32, 16, 8 }, -}; - #define RGB_TILE_WIDTH_ALIGNMENT 64 #define RGB_TILE_HEIGHT_ALIGNMENT 16 #define UBWC_PLANE_SIZE_ALIGNMENT 4096 -static const struct tile_alignment * -fdl6_tile_alignment(struct fdl_layout *layout) +static bool +is_r8g8(struct fdl_layout *layout) { - debug_assert(layout->cpp < ARRAY_SIZE(tile_alignment)); + return layout->cpp == 2 && + util_format_get_nr_components(layout->format) == 2; +} - if ((layout->cpp == 2) && (util_format_get_nr_components(layout->format) == 2)) - return &tile_alignment[0]; - else - return &tile_alignment[layout->cpp]; +void +fdl6_get_ubwc_blockwidth(struct fdl_layout *layout, + uint32_t *blockwidth, uint32_t *blockheight) +{ + static const struct { + uint8_t width; + uint8_t height; + } blocksize[] = { + { 16, 4 }, /* cpp = 1 */ + { 16, 4 }, /* cpp = 2 */ + { 16, 4 }, /* cpp = 4 */ + { 8, 4, }, /* cpp = 8 */ + { 4, 4, }, /* cpp = 16 */ + { 4, 2 }, /* cpp = 32 */ + { 0, 0 }, /* cpp = 64 (TODO) */ + }; + + /* special case for r8g8: */ + if (is_r8g8(layout)) { + *blockwidth = 16; + *blockheight = 8; + return; + } + + uint32_t cpp = fdl_cpp_shift(layout); + assert(cpp < ARRAY_SIZE(blocksize)); + *blockwidth = blocksize[cpp].width; + *blockheight = blocksize[cpp].height; } -static int -fdl6_pitchalign(struct fdl_layout *layout, int level) +static void +fdl6_tile_alignment(struct fdl_layout *layout, uint32_t *heightalign) { - uint32_t pitchalign = 64; - if (fdl_tile_mode(layout, level)) - pitchalign = fdl6_tile_alignment(layout)->pitchalign; + layout->pitchalign = fdl_cpp_shift(layout); + *heightalign = 16; + + if (is_r8g8(layout) || layout->cpp == 1) { + layout->pitchalign = 1; + *heightalign = 32; + } else if (layout->cpp == 2) { + layout->pitchalign = 2; + } - return pitchalign; + /* note: this base_align is *probably* not always right, + * it doesn't really get tested. for example with UBWC we might + * want 4k alignment, since we align UBWC levels to 4k + */ + if (layout->cpp == 1) + layout->base_align = 64; + else if (layout->cpp == 2) + layout->base_align = 128; + else + layout->base_align = 256; } /* NOTE: good way to test this is: (for example) @@ -97,8 +105,9 @@ fdl6_layout(struct fdl_layout *layout, uint32_t mip_levels, uint32_t array_size, bool is_3d, struct fdl_slice *plane_layout) { - uint32_t offset; - uint32_t pitch0; + uint32_t offset, pitch0; + uint32_t pitchalign, heightalign; + uint32_t ubwc_blockwidth, ubwc_blockheight; assert(nr_samples > 0); layout->width0 = width0; @@ -113,37 +122,54 @@ fdl6_layout(struct fdl_layout *layout, layout->nr_samples = nr_samples; layout->layer_first = !is_3d; - if (depth0 > 1) - layout->ubwc = false; - if (tile_alignment[layout->cpp].ubwc_blockwidth == 0) - layout->ubwc = false; + fdl6_get_ubwc_blockwidth(layout, &ubwc_blockwidth, &ubwc_blockheight); - const struct tile_alignment *ta = fdl6_tile_alignment(layout); + if (depth0 > 1 || ubwc_blockwidth == 0) + layout->ubwc = false; /* in layer_first layout, the level (slice) contains just one * layer (since in fact the layer contains the slices) */ uint32_t layers_in_level = layout->layer_first ? 1 : array_size; - debug_assert(ta->pitchalign); - + /* note: for tiled+noubwc layouts, we can use a lower pitchalign + * which will affect the linear levels only, (the hardware will still + * expect the tiled alignment on the tiled levels) + */ if (layout->tile_mode) { - layout->base_align = ta->basealign; + fdl6_tile_alignment(layout, &heightalign); } else { layout->base_align = 64; + layout->pitchalign = 0; + /* align pitch to at least 16 pixels: + * both turnip and galium assume there is enough alignment for 16x4 + * aligned gmem store. turnip can use CP_BLIT to work without this + * extra alignment, but gallium driver doesn't implement it yet + */ + if (layout->cpp > 4) + layout->pitchalign = fdl_cpp_shift(layout) - 2; + + /* when possible, use a bit more alignment than necessary + * presumably this is better for performance? + */ + if (!plane_layout) + layout->pitchalign = fdl_cpp_shift(layout); + + /* not used, avoid "may be used uninitialized" warning */ + heightalign = 1; } + pitchalign = 64 << layout->pitchalign; + if (plane_layout) { offset = plane_layout->offset; pitch0 = plane_layout->pitch; - if (align(pitch0, fdl6_pitchalign(layout, 0) * layout->cpp) != pitch0) - return false; - pitch0 /= layout->cpp; /* explicit pitch is in bytes */ - if (pitch0 < width0 && height0 > 1) + if (align(pitch0, pitchalign) != pitch0) return false; } else { + uint32_t nblocksx = util_format_get_nblocksx(format, width0); offset = 0; - pitch0 = util_align_npot(width0, fdl6_pitchalign(layout, 0)); + pitch0 = util_align_npot(nblocksx * layout->cpp, pitchalign); } uint32_t ubwc_width0 = width0; @@ -159,15 +185,11 @@ fdl6_layout(struct fdl_layout *layout, ubwc_height0 = util_next_power_of_two(height0); ubwc_tile_height_alignment = 64; } - ubwc_width0 = align(DIV_ROUND_UP(ubwc_width0, ta->ubwc_blockwidth), + ubwc_width0 = align(DIV_ROUND_UP(ubwc_width0, ubwc_blockwidth), RGB_TILE_WIDTH_ALIGNMENT); - ubwc_height0 = align(DIV_ROUND_UP(ubwc_height0, - ta->ubwc_blockheight), + ubwc_height0 = align(DIV_ROUND_UP(ubwc_height0, ubwc_blockheight), ubwc_tile_height_alignment); - layout->pitchalign = - util_logbase2_ceil(fdl6_pitchalign(layout, mip_levels - 1) * layout->cpp >> 6); - for (uint32_t level = 0; level < mip_levels; level++) { uint32_t depth = u_minify(depth0, level); struct fdl_slice *slice = &layout->slices[level]; @@ -184,7 +206,7 @@ fdl6_layout(struct fdl_layout *layout, uint32_t nblocksy = util_format_get_nblocksy(format, height); if (tile_mode) - nblocksy = align(nblocksy, ta->heightalign); + nblocksy = align(nblocksy, heightalign); /* The blits used for mem<->gmem work at a granularity of * 16x4, which can cause faults due to over-fetch on the @@ -196,14 +218,8 @@ fdl6_layout(struct fdl_layout *layout, if (level == mip_levels - 1) height = align(nblocksy, 4); - uint32_t nblocksx = - util_align_npot(util_format_get_nblocksx(format, u_minify(pitch0, level)), - fdl6_pitchalign(layout, level)); - - slice->offset = offset + layout->size; - uint32_t blocks = nblocksx * nblocksy; - - slice->pitch = nblocksx * layout->cpp; + slice->offset = layout->size; + slice->pitch = align(u_minify(pitch0, level), pitchalign); /* 1d array and 2d array textures must all have the same layer size * for each miplevel on a6xx. 3d textures can have different layer @@ -213,12 +229,12 @@ fdl6_layout(struct fdl_layout *layout, */ if (is_3d) { if (level < 1 || layout->slices[level - 1].size0 > 0xf000) { - slice->size0 = align(blocks * layout->cpp, 4096); + slice->size0 = align(nblocksy * slice->pitch, 4096); } else { slice->size0 = layout->slices[level - 1].size0; } } else { - slice->size0 = blocks * layout->cpp; + slice->size0 = nblocksy * slice->pitch; } layout->size += slice->size0 * depth * layers_in_level; @@ -260,12 +276,3 @@ fdl6_layout(struct fdl_layout *layout, return true; } - -void -fdl6_get_ubwc_blockwidth(struct fdl_layout *layout, - uint32_t *blockwidth, uint32_t *blockheight) -{ - const struct tile_alignment *ta = fdl6_tile_alignment(layout); - *blockwidth = ta->ubwc_blockwidth; - *blockheight = ta->ubwc_blockheight; -} diff --git a/src/freedreno/fdl/fd6_layout_test.c b/src/freedreno/fdl/fd6_layout_test.c index c5b693a931d..2a8083a866c 100644 --- a/src/freedreno/fdl/fd6_layout_test.c +++ b/src/freedreno/fdl/fd6_layout_test.c @@ -356,10 +356,10 @@ static const struct testcase testcases[] = { { .offset = 8192, .pitch = 128 }, { .offset = 12288, .pitch = 128 }, { .offset = 16384, .pitch = 128 }, - { .offset = 20480, .pitch = 64 }, - { .offset = 20544, .pitch = 64 }, - { .offset = 20608, .pitch = 64 }, - { .offset = 20672, .pitch = 64 }, + { .offset = 20480, .pitch = 128 }, + { .offset = 20608, .pitch = 128 }, + { .offset = 20736, .pitch = 128 }, + { .offset = 20864, .pitch = 128 }, }, }, }, diff --git a/src/freedreno/vulkan/tu_clear_blit.c b/src/freedreno/vulkan/tu_clear_blit.c index 0ebe3ba2c90..df2359f3bb0 100644 --- a/src/freedreno/vulkan/tu_clear_blit.c +++ b/src/freedreno/vulkan/tu_clear_blit.c @@ -1129,10 +1129,6 @@ tu_copy_buffer_to_image(struct tu_cmd_buffer *cmd, uint32_t pitch = src_width * vk_format_get_blocksize(src_format); uint32_t layer_size = src_height * pitch; - /* note: the src_va/pitch alignment of 64 is for 2D engine, - * it is also valid for 1cpp format with shader path (stencil aspect path) - */ - ops->setup(cmd, cs, dst_format, ROTATE_0, false, mask); struct tu_image_view dst; @@ -1212,10 +1208,6 @@ tu_copy_image_to_buffer(struct tu_cmd_buffer *cmd, uint32_t pitch = dst_width * vk_format_get_blocksize(dst_format); uint32_t layer_size = pitch * dst_height; - /* note: the dst_va/pitch alignment of 64 is for 2D engine, - * it is also valid for 1cpp format with shader path (stencil aspect) - */ - ops->setup(cmd, cs, dst_format, ROTATE_0, false, 0xf); struct tu_image_view src; diff --git a/src/gallium/drivers/freedreno/freedreno_resource.c b/src/gallium/drivers/freedreno/freedreno_resource.c index f7111a499e7..7343dbcc9c6 100644 --- a/src/gallium/drivers/freedreno/freedreno_resource.c +++ b/src/gallium/drivers/freedreno/freedreno_resource.c @@ -1032,9 +1032,11 @@ fd_resource_from_handle(struct pipe_screen *pscreen, uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw * rsc->layout.cpp; - /* use 64 pitchalign on a6xx where gmem_alignw is not right */ + /* pitchalign is 64-bytes for linear formats on a6xx + * layout_resource_for_modifier will validate tiled pitch + */ if (is_a6xx(screen)) - pitchalign = 64 * rsc->layout.cpp; + pitchalign = 64; if ((slice->pitch < align(prsc->width0 * rsc->layout.cpp, pitchalign)) || (slice->pitch & (pitchalign - 1)))