KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

gallium: add pipe_grid_info::last_block 

The OpenMAX state tracker will use this.

RadeonSI is adapted to use pipe_grid_info::last_block instead of its
internal state.

Acked-by: Leo Liu <leo.liu@amd.com>
This commit is contained in:
Marek Olšák 2019-02-27 17:19:53 -05:00 committed by Leo Liu
parent 34b3b92bbe
commit b9e02fe138
8 changed files with 35 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
src/gallium/auxiliary/util/u_screen.c
View File

@ -341,6 +341,9 @@ u_pipe_screen_get_param_defaults(struct pipe_screen *pscreen,
case PIPE_CAP_MAX_VARYINGS:
return 8;
case PIPE_CAP_COMPUTE_GRID_INFO_LAST_BLOCK:
return 0;
default:
unreachable("bad PIPE_CAP_*");
}

2
src/gallium/docs/source/screen.rst
View File

@ -492,6 +492,8 @@ The integer capabilities:
varyings. This will generally correspond to
``PIPE_SHADER_CAP_MAX_INPUTS`` for the fragment shader, but in some
cases may be a smaller number.
* ``PIPE_CAP_COMPUTE_GRID_INFO_LAST_BLOCK``: Whether pipe_grid_info::last_block
is implemented by the driver. See struct pipe_grid_info for more details.
.. _pipe_capf:

2
src/gallium/drivers/radeonsi/si_compute.c
View File

@ -804,7 +804,7 @@ static void si_emit_dispatch_packets(struct si_context *sctx,
* allow launching waves out-of-order. (same as Vulkan) */
S_00B800_ORDER_MODE(sctx->chip_class >= CIK);
uint *last_block = sctx->compute_last_block;
uint *last_block = info->last_block;
bool partial_block_en = last_block[0] || last_block[1] || last_block[2];
radeon_set_sh_reg_seq(cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);

18
src/gallium/drivers/radeonsi/si_compute_blit.c
View File

@ -381,7 +381,7 @@ void si_compute_copy_image(struct si_context *sctx,
si_create_copy_image_compute_shader_1d_array(ctx);
ctx->bind_compute_state(ctx, sctx->cs_copy_image_1d_array);
info.block[0] = 64;
sctx->compute_last_block[0] = width % 64;
info.last_block[0] = width % 64;
info.block[1] = 1;
info.block[2] = 1;
info.grid[0] = DIV_ROUND_UP(width, 64);
@ -392,9 +392,9 @@ void si_compute_copy_image(struct si_context *sctx,
sctx->cs_copy_image = si_create_copy_image_compute_shader(ctx);
ctx->bind_compute_state(ctx, sctx->cs_copy_image);
info.block[0] = 8;
sctx->compute_last_block[0] = width % 8;
info.last_block[0] = width % 8;
info.block[1] = 8;
sctx->compute_last_block[1] = height % 8;
info.last_block[1] = height % 8;
info.block[2] = 1;
info.grid[0] = DIV_ROUND_UP(width, 8);
info.grid[1] = DIV_ROUND_UP(height, 8);
@ -403,9 +403,6 @@ void si_compute_copy_image(struct si_context *sctx,
ctx->launch_grid(ctx, &info);
sctx->compute_last_block[0] = 0;
sctx->compute_last_block[1] = 0;
sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH |
(sctx->chip_class <= VI ? SI_CONTEXT_WRITEBACK_GLOBAL_L2 : 0) |
si_get_flush_flags(sctx, SI_COHERENCY_SHADER, L2_STREAM);
@ -483,9 +480,9 @@ void si_compute_clear_render_target(struct pipe_context *ctx,
sctx->cs_clear_render_target = si_clear_render_target_shader(ctx);
ctx->bind_compute_state(ctx, sctx->cs_clear_render_target);
info.block[0] = 8;
sctx->compute_last_block[0] = width % 8;
info.last_block[0] = width % 8;
info.block[1] = 8;
sctx->compute_last_block[1] = height % 8;
info.last_block[1] = height % 8;
info.block[2] = 1;
info.grid[0] = DIV_ROUND_UP(width, 8);
info.grid[1] = DIV_ROUND_UP(height, 8);
@ -496,7 +493,7 @@ void si_compute_clear_render_target(struct pipe_context *ctx,
si_clear_render_target_shader_1d_array(ctx);
ctx->bind_compute_state(ctx, sctx->cs_clear_render_target_1d_array);
info.block[0] = 64;
sctx->compute_last_block[0] = width % 64;
info.last_block[0] = width % 64;
info.block[1] = 1;
info.block[2] = 1;
info.grid[0] = DIV_ROUND_UP(width, 64);
@ -506,9 +503,6 @@ void si_compute_clear_render_target(struct pipe_context *ctx,
ctx->launch_grid(ctx, &info);
sctx->compute_last_block[0] = 0;
sctx->compute_last_block[1] = 0;
sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH |
(sctx->chip_class <= VI ? SI_CONTEXT_WRITEBACK_GLOBAL_L2 : 0) |
si_get_flush_flags(sctx, SI_COHERENCY_SHADER, L2_STREAM);

1
src/gallium/drivers/radeonsi/si_get.c
View File

@ -160,6 +160,7 @@ static int si_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
case PIPE_CAP_TGSI_BALLOT:
case PIPE_CAP_TGSI_VOTE:
case PIPE_CAP_TGSI_FS_FBFETCH:
case PIPE_CAP_COMPUTE_GRID_INFO_LAST_BLOCK:
return 1;
case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:

22
src/gallium/drivers/radeonsi/si_pipe.h
View File

@ -921,28 +921,6 @@ struct si_context {
uint32_t vs_blit_sh_data[SI_VS_BLIT_SGPRS_POS_TEXCOORD];
uint32_t cs_user_data[4];
/**
* last_block allows disabling threads at the farthermost grid boundary.
* Full blocks as specified by "block" are launched, but the threads
* outside of "last_block" dimensions are disabled.
*
* If a block touches the grid boundary in the i-th axis, threads with
* THREAD_ID[i] >= last_block[i] are disabled.
*
* If last_block[i] is 0, it has the same behavior as last_block[i] = block[i],
* meaning no effect.
*
* It's equivalent to doing this at the beginning of the compute shader:
*
* for (i = 0; i < 3; i++) {
* if (block_id[i] == grid[i] - 1 &&
* last_block[i] && last_block[i] >= thread_id[i])
* return;
* }
* (this could be moved into pipe_grid_info)
*/
uint compute_last_block[3];
/* Vertex and index buffers. */
bool vertex_buffers_dirty;
bool vertex_buffer_pointer_dirty;

1
src/gallium/include/pipe/p_defines.h
View File

@ -858,6 +858,7 @@ enum pipe_cap
PIPE_CAP_DEST_SURFACE_SRGB_CONTROL,
PIPE_CAP_NIR_COMPACT_ARRAYS,
PIPE_CAP_MAX_VARYINGS,
PIPE_CAP_COMPUTE_GRID_INFO_LAST_BLOCK,
};
/**

21
src/gallium/include/pipe/p_state.h
View File

@ -838,6 +838,27 @@ struct pipe_grid_info
*/
uint block[3];
/**
* last_block allows disabling threads at the farthermost grid boundary.
* Full blocks as specified by "block" are launched, but the threads
* outside of "last_block" dimensions are disabled.
*
* If a block touches the grid boundary in the i-th axis, threads with
* THREAD_ID[i] >= last_block[i] are disabled.
*
* If last_block[i] is 0, it has the same behavior as last_block[i] = block[i],
* meaning no effect.
*
* It's equivalent to doing this at the beginning of the compute shader:
*
* for (i = 0; i < 3; i++) {
* if (block_id[i] == grid[i] - 1 &&
* last_block[i] && thread_id[i] >= last_block[i])
* return;
* }
*/
uint last_block[3];
/**
* Determine the layout of the grid (in block units) to be used.
*/