[Mesa-dev] [PATCH 2/2] r600g: Pass dimension parameter to compute shader.

[Tom Stellard]

On Thu, Jul 31, 2014 at 01:28:45PM -0400, Jan Vesely wrote:
> On Thu, 2014-07-31 at 11:02 -0400, Tom Stellard wrote:
> > On Wed, Jul 30, 2014 at 07:11:35PM -0400, Jan Vesely wrote:
> > > Make the function static.
> > > 
> > 
> > No need to cc llvm-commits on these mesa patches.  Reviewers follow both
> > lists.
> 
> sorry about that
> 
> > 
> > > This needs corresponding change in LLVM otherwise it breaks parameter passing
> > > 
> > > CC: Tom Stellard <tom at stellard.net>
> > > CC: Matt Arsenault <Matthew.Arsenault at amd.com>
> > > 
> > > Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
> > > ---
> > >  src/gallium/drivers/r600/evergreen_compute.c | 26 +++++++++++++++++++++-----
> > >  src/gallium/drivers/r600/evergreen_compute.h |  1 -
> > >  2 files changed, 21 insertions(+), 6 deletions(-)
> > > 
> > > diff --git a/src/gallium/drivers/r600/evergreen_compute.c b/src/gallium/drivers/r600/evergreen_compute.c
> > > index 3928676..150bc5c 100644
> > > --- a/src/gallium/drivers/r600/evergreen_compute.c
> > > +++ b/src/gallium/drivers/r600/evergreen_compute.c
> > > @@ -266,24 +266,31 @@ static void evergreen_bind_compute_state(struct pipe_context *ctx_, void *state)
> > >   * DWORDS 3-5: Number of global work items in each dimension (x,y,z)
> > >   * DWORDS 6-8: Number of work items within each work group in each dimension
> > >   *             (x,y,z)
> > > - * DWORDS 9+ : Kernel parameters
> > > + * DWORD  9  : work dimension (needs new enough llvm)
> > > + * DWORDS 10+: Kernel parameters
> > 
> > I would prefer to add new parameters after the kernel arguments, so we
> > don't need to break compatibility with LLVM every time we add a new
> > parameter.
> 
> How would this work with multiple kernel in one module? Is it possible
> to know what kernel is the lowered code going to end up in? or did you
> have a static offset in mind?
> 

I was thinking of using a static offset from the end of the kernel
arguments.  So KernelArgSize + 0 would be number of dimensions
and it would be available to all kernels.

Then the compiler could lower any use of get_dims to a load
from the input buffer at the correct offset.

-Tom

> jan
> 
> 
> 
> > 
> > 
> > -Tom
> > 
> > >   */
> > > -void evergreen_compute_upload_input(
> > > +static void evergreen_compute_upload_input(
> > >  	struct pipe_context *ctx_,
> > >  	const uint *block_layout,
> > >  	const uint *grid_layout,
> > > -	const void *input)
> > > +	const void *input,
> > > +	unsigned dimensions)
> > >  {
> > >  	struct r600_context *ctx = (struct r600_context *)ctx_;
> > >  	struct r600_pipe_compute *shader = ctx->cs_shader_state.shader;
> > >  	unsigned i;
> > > -	/* We need to reserve 9 dwords (36 bytes) for implicit kernel
> > > +	/* We need to reserve 10 dwords (40 bytes) for implicit kernel
> > >  	 * parameters.
> > >  	 */
> > > +#if HAVE_LLVM >= 0x0306
> > > +	unsigned input_size = shader->input_size + 40;
> > > +#else
> > >  	unsigned input_size = shader->input_size + 36;
> > > +#endif
> > >  	uint32_t * num_work_groups_start;
> > >  	uint32_t * global_size_start;
> > >  	uint32_t * local_size_start;
> > > +	uint32_t * work_dim;
> > >  	uint32_t * kernel_parameters_start;
> > >  	struct pipe_box box;
> > >  	struct pipe_transfer *transfer = NULL;
> > > @@ -306,7 +313,14 @@ void evergreen_compute_upload_input(
> > >  			&box, &transfer);
> > >  	global_size_start = num_work_groups_start + (3 * (sizeof(uint) /4));
> > >  	local_size_start = global_size_start + (3 * (sizeof(uint)) / 4);
> > > +#if HAVE_LLVM >= 0x0306
> > > +	work_dim = local_size_start + (3 * (sizeof(uint)) / 4);
> > > +	work_dim[0] = dimensions;
> > > +
> > > +	kernel_parameters_start = work_dim + (1 * (sizeof(uint)) / 4);
> > > +#else
> > >  	kernel_parameters_start = local_size_start + (3 * (sizeof(uint)) / 4);
> > > +#endif
> > >  
> > >  	/* Copy the work group size */
> > >  	memcpy(num_work_groups_start, grid_layout, 3 * sizeof(uint));
> > > @@ -319,6 +333,7 @@ void evergreen_compute_upload_input(
> > >  	/* Copy the local dimensions */
> > >  	memcpy(local_size_start, block_layout, 3 * sizeof(uint));
> > >  
> > > +
> > >  	/* Copy the kernel inputs */
> > >  	memcpy(kernel_parameters_start, input, shader->input_size);
> > >  
> > > @@ -584,7 +599,8 @@ static void evergreen_launch_grid(
> > >  #endif
> > >  	shader->active_kernel = kernel;
> > >  	ctx->cs_shader_state.kernel_index = pc;
> > > -	evergreen_compute_upload_input(ctx_, block_layout, grid_layout, input);
> > > +	evergreen_compute_upload_input(ctx_, block_layout, grid_layout, input,
> > > +	                               dimensions);
> > >  	compute_emit_cs(ctx, block_layout, grid_layout);
> > >  }
> > >  
> > > diff --git a/src/gallium/drivers/r600/evergreen_compute.h b/src/gallium/drivers/r600/evergreen_compute.h
> > > index 4fb53a1..570ab2a 100644
> > > --- a/src/gallium/drivers/r600/evergreen_compute.h
> > > +++ b/src/gallium/drivers/r600/evergreen_compute.h
> > > @@ -40,7 +40,6 @@ struct r600_resource_global {
> > >  
> > >  void *evergreen_create_compute_state(struct pipe_context *ctx, const struct pipe_compute_state *cso);
> > >  void evergreen_delete_compute_state(struct pipe_context *ctx, void *state);
> > > -void evergreen_compute_upload_input(struct pipe_context *context, const uint *block_layout, const uint *grid_layout, const void *input);
> > >  void evergreen_init_atom_start_compute_cs(struct r600_context *rctx);
> > >  void evergreen_init_compute_state_functions(struct r600_context *rctx);
> > >  void evergreen_emit_cs_shader(struct r600_context *rctx, struct r600_atom * atom);
> > > -- 
> > > 1.9.3
> > > 
> 
> -- 
> Jan Vesely <jan.vesely at rutgers.edu>