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 <t...@stellard.net> > > > CC: Matt Arsenault <matthew.arsena...@amd.com> > > > > > > Signed-off-by: Jan Vesely <jan.ves...@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.ves...@rutgers.edu> _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev