Skip to content

Commit

Permalink
drm/amd/display: Update MALL SS NumWays calculation
Browse files Browse the repository at this point in the history 
[Description]
Update MALL SS NumWays calculation according
to programming guide.

Reviewed-by: Jun Lei <Jun.Lei@amd.com>
Acked-by: Jasdeep Dhillon <jdhillon@amd.com>
Signed-off-by: Alvin Lee <Alvin.Lee2@amd.com>
Tested-by: Daniel Wheeler <daniel.wheeler@amd.com>
  • Loading branch information
Alvin Lee authored and Asher Song committed Sep 27, 2022
1 parent 03aa9bd commit 3782687
Show file tree
Hide file tree
Showing 2 changed files with 98 additions and 110 deletions.
1 change: 1 addition & 0 deletions drivers/gpu/drm/amd/display/dc/dc.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -837,6 +837,7 @@ struct dc_debug_options {
bool force_subvp_mclk_switch;
bool allow_sw_cursor_fallback;
unsigned int force_subvp_num_ways;
unsigned int force_mall_ss_num_ways;
bool alloc_extra_way_for_cursor;
bool force_usr_allow;
/* uses value at boot and disables switch */
Expand Down
207 changes: 97 additions & 110 deletions drivers/gpu/drm/amd/display/dc/dcn32/dcn32_hwseq.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,7 @@
#include "dcn20/dcn20_optc.h"
#include "dmub_subvp_state.h"
#include "dce/dmub_hw_lock_mgr.h"
#include "dcn32_resource.h"
#include "dc_link_dp.h"
#include "dmub/inc/dmub_subvp_state.h"

Expand Down Expand Up @@ -199,139 +200,123 @@ static bool dcn32_check_no_memory_request_for_cab(struct dc *dc)
return false;
}

/* This function takes in the start address and surface size to be cached in CAB
* and calculates the total number of cache lines required to store the surface.
* The number of cache lines used for each surface is calculated independently of
* one another. For example, if there is a primary surface(1), meta surface(2), and
* cursor(3), this function should be called 3 times to calculate the number of cache
* lines used for each of those surfaces.
*/
static uint32_t dcn32_cache_lines_for_surface(struct dc *dc, uint32_t surface_size, uint64_t start_address)
{
uint32_t lines_used = 1;
uint32_t num_cached_bytes = 0;
uint32_t remaining_size = 0;
uint32_t cache_line_size = dc->caps.cache_line_size;
uint32_t remainder = 0;

/* 1. Calculate surface size minus the number of bytes stored
* in the first cache line (all bytes in first cache line might
* not be fully used).
*/
div_u64_rem(start_address, cache_line_size, &remainder);
num_cached_bytes = cache_line_size - remainder;
remaining_size = surface_size - num_cached_bytes;

/* 2. Calculate number of cache lines that will be fully used with
* the remaining number of bytes to be stored.
*/
lines_used += (remaining_size / cache_line_size);

/* 3. Check if we need an extra line due to the remaining size not being
* a multiple of CACHE_LINE_SIZE.
*/
if (remaining_size % cache_line_size > 0)
lines_used++;

return lines_used;
}

/* This function loops through every surface that needs to be cached in CAB for SS,
* and calculates the total number of ways required to store all surfaces (primary,
* meta, cursor).
*/
static uint32_t dcn32_calculate_cab_allocation(struct dc *dc, struct dc_state *ctx)
{
uint8_t i, j;
uint8_t i;
int j;
struct dc_stream_state *stream = NULL;
struct dc_plane_state *plane = NULL;
uint32_t surface_size = 0;
uint32_t cursor_size = 0;
uint32_t cache_lines_used = 0;
uint32_t total_lines = 0;
uint32_t lines_per_way = 0;
uint32_t num_ways = 0;
uint32_t prev_addr_low = 0;
uint8_t num_ways = 0;
uint8_t bytes_per_pixel = 0;
uint8_t cursor_bpp = 0;
uint16_t mblk_width = 0;
uint16_t mblk_height = 0;
uint16_t mall_alloc_width_blk_aligned = 0;
uint16_t mall_alloc_height_blk_aligned = 0;
uint16_t num_mblks = 0;
uint32_t bytes_in_mall = 0;
uint32_t cache_lines_used = 0;
uint32_t cache_lines_per_plane = 0;

for (i = 0; i < ctx->stream_count; i++) {
stream = ctx->streams[i];
for (i = 0; i < dc->res_pool->pipe_count; i++) {
struct pipe_ctx *pipe = &dc->current_state->res_ctx.pipe_ctx[i];

// Don't include PSR surface in the total surface size for CAB allocation
if (stream->link->psr_settings.psr_version != DC_PSR_VERSION_UNSUPPORTED)
if (!pipe->stream || !pipe->plane_state ||
pipe->stream->link->psr_settings.psr_version != DC_PSR_VERSION_UNSUPPORTED ||
pipe->stream->mall_stream_config.type == SUBVP_PHANTOM)
continue;

if (ctx->stream_status[i].plane_count == 0)
continue;
bytes_per_pixel = pipe->plane_state->format >= SURFACE_PIXEL_FORMAT_GRPH_ARGB16161616 ? 8 : 4;
mblk_width = DCN3_2_MBLK_WIDTH;
mblk_height = bytes_per_pixel == 4 ? DCN3_2_MBLK_HEIGHT_4BPE : DCN3_2_MBLK_HEIGHT_8BPE;

// For each stream, loop through each plane to calculate the number of cache
// lines required to store the surface in CAB
for (j = 0; j < ctx->stream_status[i].plane_count; j++) {
plane = ctx->stream_status[i].plane_states[j];
/* full_vp_width_blk_aligned = FLOOR(vp_x_start + full_vp_width + blk_width - 1, blk_width) -
* FLOOR(vp_x_start, blk_width)
*
* mall_alloc_width_blk_aligned_l/c = full_vp_width_blk_aligned_l/c
*/
mall_alloc_width_blk_aligned = ((pipe->plane_res.scl_data.viewport.x +
pipe->plane_res.scl_data.viewport.width + mblk_width - 1) / mblk_width * mblk_width) +
(pipe->plane_res.scl_data.viewport.x / mblk_width * mblk_width);

/* full_vp_height_blk_aligned = FLOOR(vp_y_start + full_vp_height + blk_height - 1, blk_height) -
* FLOOR(vp_y_start, blk_height)
*
* mall_alloc_height_blk_aligned_l/c = full_vp_height_blk_aligned_l/c
*/
mall_alloc_height_blk_aligned = ((pipe->plane_res.scl_data.viewport.y +
pipe->plane_res.scl_data.viewport.height + mblk_height - 1) / mblk_height * mblk_height) +
(pipe->plane_res.scl_data.viewport.y / mblk_height * mblk_height);

// Calculate total surface size
if (prev_addr_low != plane->address.grph.addr.u.low_part) {
/* if plane address are different from prev FB, then userspace allocated separate FBs*/
surface_size += plane->plane_size.surface_pitch *
plane->plane_size.surface_size.height *
(plane->format >= SURFACE_PIXEL_FORMAT_GRPH_ARGB16161616 ? 8 : 4);
num_mblks = ((mall_alloc_width_blk_aligned + mblk_width - 1) / mblk_width) *
((mall_alloc_height_blk_aligned + mblk_height - 1) / mblk_height);

prev_addr_low = plane->address.grph.addr.u.low_part;
} else {
/* We have the same fb for all the planes.
* Xorg always creates one giant fb that holds all surfaces,
* so allocating it once is sufficient.
* */
continue;
}
// Convert surface size + starting address to number of cache lines required
// (alignment accounted for)
cache_lines_used += dcn32_cache_lines_for_surface(dc, surface_size,
plane->address.grph.addr.quad_part);

if (plane->address.grph.meta_addr.quad_part) {
// Meta surface
cache_lines_used += dcn32_cache_lines_for_surface(dc, surface_size,
plane->address.grph.meta_addr.quad_part);
}
}
/* For DCC:
* meta_num_mblk = CEILING(full_mblk_width_ub_l*full_mblk_height_ub_l*Bpe/256/mblk_bytes, 1)
*/
if (pipe->plane_state->dcc.enable)
num_mblks += (mall_alloc_width_blk_aligned * mall_alloc_width_blk_aligned * bytes_per_pixel +
(256 * DCN3_2_MALL_MBLK_SIZE_BYTES) - 1) / (256 * DCN3_2_MALL_MBLK_SIZE_BYTES);

// Include cursor size for CAB allocation
for (j = 0; j < dc->res_pool->pipe_count; j++) {
struct pipe_ctx *pipe = &ctx->res_ctx.pipe_ctx[j];
struct hubp *hubp = pipe->plane_res.hubp;
bytes_in_mall = num_mblks * DCN3_2_MALL_MBLK_SIZE_BYTES;

if (pipe->stream && pipe->plane_state && hubp)
/* Find the cursor plane and use the exact size instead of
* using the max for calculation
*/
if (hubp->curs_attr.width > 0) {
// Round cursor width to next multiple of 64
cursor_size = (((hubp->curs_attr.width + 63) / 64) * 64) * hubp->curs_attr.height;
break;
}
}
/* (cache lines used is total bytes / cache_line size. Add +2 for worst case alignment
* (MALL is 64-byte aligned)
*/
cache_lines_per_plane = bytes_in_mall / dc->caps.cache_line_size + 2;
cache_lines_used += cache_lines_per_plane;
}

switch (stream->cursor_attributes.color_format) {
case CURSOR_MODE_MONO:
cursor_size /= 2;
break;
case CURSOR_MODE_COLOR_1BIT_AND:
case CURSOR_MODE_COLOR_PRE_MULTIPLIED_ALPHA:
case CURSOR_MODE_COLOR_UN_PRE_MULTIPLIED_ALPHA:
cursor_size *= 4;
break;
// Include cursor size for CAB allocation
for (j = 0; j < dc->res_pool->pipe_count; j++) {
struct pipe_ctx *pipe = &ctx->res_ctx.pipe_ctx[j];
struct hubp *hubp = pipe->plane_res.hubp;

case CURSOR_MODE_COLOR_64BIT_FP_PRE_MULTIPLIED:
case CURSOR_MODE_COLOR_64BIT_FP_UN_PRE_MULTIPLIED:
cursor_size *= 8;
break;
}
if (pipe->stream && pipe->plane_state && hubp)
/* Find the cursor plane and use the exact size instead of
using the max for calculation */

if (stream->cursor_position.enable && !dc->debug.alloc_extra_way_for_cursor &&
cursor_size > 16384) {
cache_lines_used += dcn32_cache_lines_for_surface(dc, cursor_size,
plane->address.grph.cursor_cache_addr.quad_part);
}
if (hubp->curs_attr.width > 0) {
// Round cursor width to next multiple of 64
cursor_size = (((hubp->curs_attr.width + 63) / 64) * 64) * hubp->curs_attr.height;

switch (pipe->stream->cursor_attributes.color_format) {
case CURSOR_MODE_MONO:
cursor_size /= 2;
cursor_bpp = 4;
break;
case CURSOR_MODE_COLOR_1BIT_AND:
case CURSOR_MODE_COLOR_PRE_MULTIPLIED_ALPHA:
case CURSOR_MODE_COLOR_UN_PRE_MULTIPLIED_ALPHA:
cursor_size *= 4;
cursor_bpp = 4;
break;

case CURSOR_MODE_COLOR_64BIT_FP_PRE_MULTIPLIED:
case CURSOR_MODE_COLOR_64BIT_FP_UN_PRE_MULTIPLIED:
cursor_size *= 8;
cursor_bpp = 8;
break;
}

if (pipe->stream->cursor_position.enable && !dc->debug.alloc_extra_way_for_cursor &&
cursor_size > 16384) {
/* cursor_num_mblk = CEILING(num_cursors*cursor_width*cursor_width*cursor_Bpe/mblk_bytes, 1)
*/
cache_lines_used += (((hubp->curs_attr.width * hubp->curs_attr.height * cursor_bpp +
DCN3_2_MALL_MBLK_SIZE_BYTES - 1) / DCN3_2_MALL_MBLK_SIZE_BYTES) *
DCN3_2_MALL_MBLK_SIZE_BYTES) / dc->caps.cache_line_size + 2;
}
break;
}
}

// Convert number of cache lines required to number of ways
Expand Down Expand Up @@ -361,7 +346,9 @@ static uint32_t dcn32_calculate_cab_allocation(struct dc *dc, struct dc_state *c
}
}
}

if (dc->debug.force_mall_ss_num_ways > 0) {
num_ways = dc->debug.force_mall_ss_num_ways;
}
return num_ways;
}

Expand Down

0 comments on commit 3782687

Please sign in to comment.