[Mesa-dev] [PATCH 7/7] softpipe: add support for compute shaders.
Roland Scheidegger
sroland at vmware.com
Tue Apr 26 23:08:16 UTC 2016
Am 26.04.2016 um 23:18 schrieb Dave Airlie:
> On 27 April 2016 at 06:07, Roland Scheidegger <sroland at vmware.com> wrote:
>> Am 26.04.2016 um 06:42 schrieb Dave Airlie:
>>> From: Dave Airlie <airlied at redhat.com>
>>>
>>> This enables ARB_compute_shader on softpipe. I've only
>>> tested this with piglit so far, and I hopefully plan
>>> on integrating it with my vulkan work. I'll get to
>>> testing it with deqp more later.
>>>
>>> The basic premise is to create up to 1024 restartable
>>> TGSI machines, and execute workgroups of those machines.
>>>
>>> Signed-off-by: Dave Airlie <airlied at redhat.com>
>>> ---
>>> src/gallium/drivers/softpipe/Makefile.sources | 1 +
>>> src/gallium/drivers/softpipe/sp_compute.c | 211 +++++++++++++++++++++++++
>>> src/gallium/drivers/softpipe/sp_context.c | 3 +
>>> src/gallium/drivers/softpipe/sp_context.h | 4 +-
>>> src/gallium/drivers/softpipe/sp_screen.c | 48 +++++-
>>> src/gallium/drivers/softpipe/sp_state.h | 9 ++
>>> src/gallium/drivers/softpipe/sp_state_shader.c | 51 ++++++
>>> 7 files changed, 324 insertions(+), 3 deletions(-)
>>> create mode 100644 src/gallium/drivers/softpipe/sp_compute.c
>>>
>>> diff --git a/src/gallium/drivers/softpipe/Makefile.sources b/src/gallium/drivers/softpipe/Makefile.sources
>>> index 1d42351..d72266f 100644
>>> --- a/src/gallium/drivers/softpipe/Makefile.sources
>>> +++ b/src/gallium/drivers/softpipe/Makefile.sources
>>> @@ -4,6 +4,7 @@ C_SOURCES := \
>>> sp_clear.h \
>>> sp_context.c \
>>> sp_context.h \
>>> + sp_compute.c \
>>> sp_draw_arrays.c \
>>> sp_fence.c \
>>> sp_fence.h \
>>> diff --git a/src/gallium/drivers/softpipe/sp_compute.c b/src/gallium/drivers/softpipe/sp_compute.c
>>> new file mode 100644
>>> index 0000000..7467686
>>> --- /dev/null
>>> +++ b/src/gallium/drivers/softpipe/sp_compute.c
>>> @@ -0,0 +1,211 @@
>>> +#include "util/u_inlines.h"
>>> +#include "util/u_math.h"
>>> +#include "util/u_memory.h"
>>> +#include "util/u_pstipple.h"
>>> +#include "pipe/p_shader_tokens.h"
>>> +#include "draw/draw_context.h"
>>> +#include "draw/draw_vertex.h"
>>> +#include "sp_context.h"
>>> +#include "sp_screen.h"
>>> +#include "sp_state.h"
>>> +#include "sp_texture.h"
>>> +#include "sp_tex_sample.h"
>>> +#include "sp_tex_tile_cache.h"
>>> +#include "tgsi/tgsi_parse.h"
>>> +
>>> +static void
>>> +cs_prepare(const struct sp_compute_shader *cs,
>>> + struct tgsi_exec_machine *machine,
>>> + int w, int h, int d,
>>> + int g_w, int g_h, int g_d,
>>> + int b_w, int b_h, int b_d,
>>> + struct tgsi_sampler *sampler,
>>> + struct tgsi_image *image,
>>> + struct tgsi_buffer *buffer )
>>> +{
>>> + int j;
>>> + /*
>>> + * Bind tokens/shader to the interpreter's machine state.
>>> + */
>>> + tgsi_exec_machine_bind_shader(machine,
>>> + cs->tokens,
>>> + sampler, image, buffer);
>>> +
>>> + if (machine->SysSemanticToIndex[TGSI_SEMANTIC_THREAD_ID] != -1) {
>>> + unsigned i = machine->SysSemanticToIndex[TGSI_SEMANTIC_THREAD_ID];
>>> + for (j = 0; j < TGSI_QUAD_SIZE; j++) {
>>> + machine->SystemValue[i].xyzw[0].i[j] = w;
>>> + machine->SystemValue[i].xyzw[1].i[j] = h;
>>> + machine->SystemValue[i].xyzw[2].i[j] = d;
>>> + }
>>> + }
>>> +
>>> + if (machine->SysSemanticToIndex[TGSI_SEMANTIC_GRID_SIZE] != -1) {
>>> + unsigned i = machine->SysSemanticToIndex[TGSI_SEMANTIC_GRID_SIZE];
>>> + for (j = 0; j < TGSI_QUAD_SIZE; j++) {
>>> + machine->SystemValue[i].xyzw[0].i[j] = g_w;
>>> + machine->SystemValue[i].xyzw[1].i[j] = g_h;
>>> + machine->SystemValue[i].xyzw[2].i[j] = g_d;
>>> + }
>>> + }
>>> +
>>> + if (machine->SysSemanticToIndex[TGSI_SEMANTIC_BLOCK_SIZE] != -1) {
>>> + unsigned i = machine->SysSemanticToIndex[TGSI_SEMANTIC_BLOCK_SIZE];
>>> + for (j = 0; j < TGSI_QUAD_SIZE; j++) {
>>> + machine->SystemValue[i].xyzw[0].i[j] = b_w;
>>> + machine->SystemValue[i].xyzw[1].i[j] = b_h;
>>> + machine->SystemValue[i].xyzw[2].i[j] = b_d;
>>> + }
>>> + }
>>> +}
>>> +
>>> +static bool
>>> +cs_run(const struct sp_compute_shader *cs,
>>> + int g_w, int g_h, int g_d,
>>> + struct tgsi_exec_machine *machine, bool restart)
>>> +{
>>> + if (!restart) {
>>> + if (machine->SysSemanticToIndex[TGSI_SEMANTIC_BLOCK_ID] != -1) {
>>> + unsigned i = machine->SysSemanticToIndex[TGSI_SEMANTIC_BLOCK_ID];
>>> + int j;
>>> + for (j = 0; j < TGSI_QUAD_SIZE; j++) {
>>> + machine->SystemValue[i].xyzw[0].i[j] = g_w;
>>> + machine->SystemValue[i].xyzw[1].i[j] = g_h;
>>> + machine->SystemValue[i].xyzw[2].i[j] = g_d;
>>> + }
>>> + }
>>> + machine->NonHelperMask = (1 << 1) - 1;
>>> + }
>>> +
>>> + tgsi_exec_machine_run(machine, restart ? machine->pc : 0);
>>> +
>>> + if (machine->pc != -1)
>>> + return true;
>>> + return false;
>>> +}
>>> +
>>> +static void
>>> +run_workgroup(const struct sp_compute_shader *cs,
>>> + int g_w, int g_h, int g_d, int num_threads,
>>> + struct tgsi_exec_machine **machines)
>>> +{
>>> + int i;
>>> + bool grp_hit_barrier, restart_threads = false;
>>> +
>>> + do {
>>> + grp_hit_barrier = false;
>>> + for (i = 0; i < num_threads; i++) {
>>> + grp_hit_barrier |= cs_run(cs, g_w, g_h, g_d, machines[i], restart_threads);
>>> + }
>>> + restart_threads = false;
>>> + if (grp_hit_barrier) {
>>> + grp_hit_barrier = false;
>>> + restart_threads = true;
>>> + }
>>> + } while (restart_threads);
>>> +}
>>> +
>>> +static void
>>> +cs_delete(const struct sp_compute_shader *cs,
>>> + struct tgsi_exec_machine *machine)
>>> +{
>>> + if (machine->Tokens == cs->tokens) {
>>> + tgsi_exec_machine_bind_shader(machine, NULL, NULL, NULL, NULL);
>>> + }
>>> +}
>>> +
>>> +static void
>>> +fill_grid_size(struct pipe_context *context,
>>> + const struct pipe_grid_info *info,
>>> + uint32_t grid_size[3])
>>> +{
>>> + struct pipe_transfer *transfer;
>>> + uint32_t *params;
>>> + if (!info->indirect) {
>>> + grid_size[0] = info->grid[0];
>>> + grid_size[1] = info->grid[1];
>>> + grid_size[2] = info->grid[2];
>>> + return;
>>> + }
>>> + params = pipe_buffer_map_range(context, info->indirect,
>>> + info->indirect_offset,
>>> + 3 * sizeof(uint32_t),
>>> + PIPE_TRANSFER_READ,
>>> + &transfer);
>>> +
>>> + if (!transfer)
>>> + return;
>>> +
>>> + grid_size[0] = params[0];
>>> + grid_size[1] = params[1];
>>> + grid_size[2] = params[2];
>>> + pipe_buffer_unmap(context, transfer);
>>> +}
>>> +
>>> +void
>>> +softpipe_launch_grid(struct pipe_context *context,
>>> + const struct pipe_grid_info *info)
>>> +{
>>> + struct softpipe_context *softpipe = softpipe_context(context);
>>> + struct sp_compute_shader *cs = softpipe->cs;
>>> + int num_threads_in_group;
>>> + struct tgsi_exec_machine **machines;
>>> + int bwidth, bheight, bdepth;
>>> + int w, h, d, i;
>>> + int g_w, g_h, g_d;
>>> + uint32_t grid_size[3];
>>> + void *local_mem = NULL;
>>> +
>>> + bwidth = cs->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH];
>>> + bheight = cs->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT];
>>> + bdepth = cs->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
>>> + num_threads_in_group = bwidth * bheight * bdepth;
>>> +
>>> + fill_grid_size(context, info, grid_size);
>>> +
>>> + if (cs->shader.req_local_mem) {
>>> + local_mem = CALLOC(1, cs->shader.req_local_mem);
>>> + }
>>> +
>>> + machines = CALLOC(sizeof(struct tgsi_exec_machine *), num_threads_in_group);
>>> + if (!machines)
>>> + return;
>>> +
>>> + /* initialise machines + GRID_SIZE + THREAD_ID + BLOCK_SIZE */
>>> + for (d = 0; d < bdepth; d++) {
>>> + for (h = 0; h < bheight; h++) {
>>> + for (w = 0; w < bwidth; w++) {
>>> + int idx = w + (h * bwidth) + (d * bheight * bwidth);
>>> + machines[idx] = tgsi_exec_machine_create(PIPE_SHADER_COMPUTE);
>>> +
>>> + machines[idx]->LocalMem = local_mem;
>>> + machines[idx]->LocalMemSize = cs->shader.req_local_mem;
>>> + cs_prepare(cs, machines[idx],
>>> + w, h, d,
>>> + grid_size[0], grid_size[1], grid_size[2],
>>> + bwidth, bheight, bdepth,
>>> + (struct tgsi_sampler *)softpipe->tgsi.sampler[PIPE_SHADER_COMPUTE],
>>> + (struct tgsi_image *)softpipe->tgsi.image[PIPE_SHADER_COMPUTE],
>>> + (struct tgsi_buffer *)softpipe->tgsi.buffer[PIPE_SHADER_COMPUTE]);
>>> + tgsi_exec_set_constant_buffers(machines[idx], PIPE_MAX_CONSTANT_BUFFERS,
>>> + softpipe->mapped_constants[PIPE_SHADER_COMPUTE],
>>> + softpipe->const_buffer_size[PIPE_SHADER_COMPUTE]);
>>> + }
>>> + }
>>> + }
>>> +
>>> + for (g_d = 0; g_d < grid_size[2]; g_d++) {
>>> + for (g_h = 0; g_h < grid_size[1]; g_h++) {
>>> + for (g_w = 0; g_w < grid_size[0]; g_w++) {
>>> + run_workgroup(cs, g_w, g_h, g_d, num_threads_in_group, machines);
>>> + }
>>> + }
>>> + }
>>> +
>>> + for (i = 0; i < num_threads_in_group; i++) {
>>> + cs_delete(cs, machines[i]);
>>> + }
>>> +
>>> + FREE(local_mem);
>>> + FREE(machines);
>>> +}
>>> diff --git a/src/gallium/drivers/softpipe/sp_context.c b/src/gallium/drivers/softpipe/sp_context.c
>>> index e3ec524..1690e38 100644
>>> --- a/src/gallium/drivers/softpipe/sp_context.c
>>> +++ b/src/gallium/drivers/softpipe/sp_context.c
>>> @@ -212,6 +212,7 @@ softpipe_create_context(struct pipe_screen *screen,
>>>
>>> softpipe->dump_fs = debug_get_bool_option( "SOFTPIPE_DUMP_FS", FALSE );
>>> softpipe->dump_gs = debug_get_bool_option( "SOFTPIPE_DUMP_GS", FALSE );
>>> + softpipe->dump_cs = debug_get_bool_option( "SOFTPIPE_DUMP_CS", FALSE );
>>>
>>> softpipe->pipe.screen = screen;
>>> softpipe->pipe.destroy = softpipe_destroy;
>>> @@ -233,6 +234,8 @@ softpipe_create_context(struct pipe_screen *screen,
>>>
>>> softpipe->pipe.draw_vbo = softpipe_draw_vbo;
>>>
>>> + softpipe->pipe.launch_grid = softpipe_launch_grid;
>>> +
>>> softpipe->pipe.clear = softpipe_clear;
>>> softpipe->pipe.flush = softpipe_flush_wrapped;
>>> softpipe->pipe.texture_barrier = softpipe_texture_barrier;
>>> diff --git a/src/gallium/drivers/softpipe/sp_context.h b/src/gallium/drivers/softpipe/sp_context.h
>>> index 70d00c8..a57f587 100644
>>> --- a/src/gallium/drivers/softpipe/sp_context.h
>>> +++ b/src/gallium/drivers/softpipe/sp_context.h
>>> @@ -71,6 +71,7 @@ struct softpipe_context {
>>> struct sp_geometry_shader *gs;
>>> struct sp_velems_state *velems;
>>> struct sp_so_state *so;
>>> + struct sp_compute_shader *cs;
>>>
>>> /** Other rendering state */
>>> struct pipe_blend_color blend_color;
>>> @@ -205,10 +206,11 @@ struct softpipe_context {
>>> * XXX wouldn't it make more sense for the tile cache to just be part
>>> * of sp_sampler_view?
>>> */
>>> - struct softpipe_tex_tile_cache *tex_cache[PIPE_SHADER_GEOMETRY+1][PIPE_MAX_SHADER_SAMPLER_VIEWS];
>>> + struct softpipe_tex_tile_cache *tex_cache[PIPE_SHADER_TYPES][PIPE_MAX_SHADER_SAMPLER_VIEWS];
>>>
>>> unsigned dump_fs : 1;
>>> unsigned dump_gs : 1;
>>> + unsigned dump_cs : 1;
>>> unsigned no_rast : 1;
>>> };
>>>
>>> diff --git a/src/gallium/drivers/softpipe/sp_screen.c b/src/gallium/drivers/softpipe/sp_screen.c
>>> index d89d95c..4beeb80 100644
>>> --- a/src/gallium/drivers/softpipe/sp_screen.c
>>> +++ b/src/gallium/drivers/softpipe/sp_screen.c
>>> @@ -157,7 +157,7 @@ softpipe_get_param(struct pipe_screen *screen, enum pipe_cap param)
>>> case PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION:
>>> return 0;
>>> case PIPE_CAP_COMPUTE:
>>> - return 0;
>>> + return 1;
>>> case PIPE_CAP_USER_VERTEX_BUFFERS:
>>> case PIPE_CAP_USER_INDEX_BUFFERS:
>>> case PIPE_CAP_USER_CONSTANT_BUFFERS:
>>> @@ -289,6 +289,8 @@ softpipe_get_shader_param(struct pipe_screen *screen, unsigned shader, enum pipe
>>> {
>>> case PIPE_SHADER_FRAGMENT:
>>> return tgsi_exec_get_shader_param(param);
>>> + case PIPE_SHADER_COMPUTE:
>>> + return tgsi_exec_get_shader_param(param);
>>> case PIPE_SHADER_VERTEX:
>>> case PIPE_SHADER_GEOMETRY:
>>> if (sp_screen->use_llvm)
>>> @@ -447,6 +449,48 @@ softpipe_get_timestamp(struct pipe_screen *_screen)
>>> return os_time_get_nano();
>>> }
>>>
>>> +static int
>>> +softpipe_get_compute_param(struct pipe_screen *_screen,
>>> + enum pipe_shader_ir ir_type,
>>> + enum pipe_compute_cap param,
>>> + void *ret)
>>> +{
>>> + switch (param) {
>>> + case PIPE_COMPUTE_CAP_IR_TARGET:
>>> + return 0;
>>> + case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
>>> + if (ret) {
>>> + uint64_t *grid_size = ret;
>>> + grid_size[0] = 65535;
>>> + grid_size[1] = 65535;
>>> + grid_size[2] = 65535;
>>> + }
>>> + return 3 * sizeof(uint64_t) ;
>>> + case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
>>> + if (ret) {
>>> + uint64_t *block_size = ret;
>>> + block_size[0] = 1024;
>>> + block_size[1] = 1024;
>>> + block_size[2] = 1024;
>>> + }
>>> + return 3 * sizeof(uint64_t);
>>> + case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
>>> + if (ret) {
>>> + uint64_t *max_threads_per_block = ret;
>>> + *max_threads_per_block = 2048;
>>> + }
>>> + return sizeof(uint64_t);
>>> + case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
>>> + if (ret) {
>>> + uint64_t *max_local_size = ret;
>>> + /* Value reported by the closed source driver. */
>>
>> The comment here doesn't make much sense...
>>
>> 1024 interpreted tgsi machines, all running serially - I'm sure
>> performance is going to be amazing.
>>
>> But the approach looks reasonable to me.
>>
>> I'm not really familiar with compute shaders, but what I'm wondering is
>> since tgsi exec always operates on 4 values at a time, is that somehow
>> implicit in compute shaders?
>
> So far I've set the execmask to 1 active channel, I'm contemplating
> changing that
> though and using less machines.
Ah yes, I think that would indeed be desirable.
>
> Any ideas how to implement this in llvm? :-) 1024 CPU threads?
I suppose 1024 is really the minimum work size you have to support?
But since things are always run 4-wide (or 8-wide) that would "only" be
256 (or 128) threads. That many threads sound a bit suboptimal to me
(unless you really have a boatload of cpu cores), but why not - I
suppose you can always pause some of the threads, not all need to be
active at the same time.
Though I wonder what the opencl-on-cpu guys do...
Roland
More information about the mesa-dev
mailing list