[PATCH 5/5] wined3d: Implement UAV clears on the Vulkan backend.
Henri Verbeet
hverbeet at gmail.com
Tue Aug 10 13:51:21 CDT 2021
On Tue, 10 Aug 2021 at 10:35, Jan Sikorski <jsikorski at codeweavers.com> wrote:
> diff --git a/dlls/wined3d/uav_clear_shaders.inc.c b/dlls/wined3d/uav_clear_shaders.inc.c
> new file mode 100644
> index 00000000000..6cb3c808578
> --- /dev/null
> +++ b/dlls/wined3d/uav_clear_shaders.inc.c
> @@ -0,0 +1,365 @@
> +static const uint32_t cs_uav_clear_buffer_float_code[] =
> +{
> +#if 0
> + RWBuffer<float4> dst;
> +
> + struct
> + {
> + float4 clear_value;
> + int2 dst_offset;
> + int2 dst_extent;
> + } u_info;
> +
> + [numthreads(128, 1, 1)]
> + void main(int3 thread_id : SV_DispatchThreadID)
> + {
> + if (thread_id.x < u_info.dst_extent.x)
> + dst[u_info.dst_offset.x + thread_id.x] = u_info.clear_value;
> + }
> +#endif
> + 0x43425844, 0xe114ba61, 0xff6a0d0b, 0x7b25c8f4, 0xfcf7cf22, 0x00000001, 0x0000010c, 0x00000003,
> + 0x0000002c, 0x0000003c, 0x0000004c, 0x4e475349, 0x00000008, 0x00000000, 0x00000008, 0x4e47534f,
> + 0x00000008, 0x00000000, 0x00000008, 0x58454853, 0x000000b8, 0x00050050, 0x0000002e, 0x0100086a,
> + 0x04000059, 0x00208e46, 0x00000000, 0x00000002, 0x0400089c, 0x0011e000, 0x00000000, 0x00005555,
> + 0x0200005f, 0x00020012, 0x02000068, 0x00000001, 0x0400009b, 0x00000080, 0x00000001, 0x00000001,
> + 0x07000022, 0x00100012, 0x00000000, 0x0002000a, 0x0020802a, 0x00000000, 0x00000001, 0x0304001f,
> + 0x0010000a, 0x00000000, 0x0700001e, 0x00100012, 0x00000000, 0x0002000a, 0x0020800a, 0x00000000,
> + 0x00000001, 0x080000a4, 0x0011e0f2, 0x00000000, 0x00100006, 0x00000000, 0x00208e46, 0x00000000,
> + 0x00000000, 0x01000015, 0x0100003e,
> +};
> +
We're missing a license header here. I don't see much point in making
this a .c file; a .h seams preferable, and if we're following the
vkd3d convention, that would be wined3d_shaders.h.
> +static void STDMETHODCALLTYPE wined3d_uav_clear_object_destroyed(void *parent)
> +{
> +}
> +
> +static struct wined3d_parent_ops wined3d_uav_clear_ops =
> +{
> + wined3d_uav_clear_object_destroyed
> +};
> +
That's "wined3d_null_parent_ops".
> +static bool create_shader(struct wined3d_device *device, const uint32_t *byte_code, size_t byte_code_size,
> + struct wined3d_shader **shader)
> +{
> + struct wined3d_shader_desc shader_desc;
> + HRESULT result;
> +
> + shader_desc.byte_code = byte_code;
> + shader_desc.byte_code_size = byte_code_size;
> +
> + result = wined3d_shader_create_cs(device, &shader_desc, NULL, &wined3d_uav_clear_ops, shader);
> + if (FAILED(result))
> + WARN("Failed to initialize shader: %#x\n", result);
> +
> + return SUCCEEDED(result);
> +}
> +
> +#include "uav_clear_shaders.inc.c"
> +
> +void wined3d_device_vk_uav_clear_state_init(struct wined3d_device_vk *device_vk)
> +{
> + struct wined3d_context_vk *context_vk = &device_vk->context_vk;
> + struct wined3d_device *device = &device_vk->d;
> + struct wined3d_uav_clear_state_vk *state = &device_vk->uav_clear_state;
> +
> + create_shader(device, cs_uav_clear_buffer_float_code, sizeof(cs_uav_clear_buffer_float_code),
> + &state->float_shaders.buffer);
> + create_shader(device, cs_uav_clear_buffer_uint_code, sizeof(cs_uav_clear_buffer_uint_code),
> + &state->uint_shaders.buffer);
> + create_shader(device, cs_uav_clear_1d_array_float_code, sizeof(cs_uav_clear_1d_array_float_code),
> + &state->float_shaders.image_1d);
> + create_shader(device, cs_uav_clear_1d_array_uint_code, sizeof(cs_uav_clear_1d_array_uint_code),
> + &state->uint_shaders.image_1d);
> + create_shader(device, cs_uav_clear_1d_float_code, sizeof(cs_uav_clear_1d_float_code),
> + &state->float_shaders.image_1d_array);
> + create_shader(device, cs_uav_clear_1d_uint_code, sizeof(cs_uav_clear_1d_uint_code),
> + &state->uint_shaders.image_1d_array);
> + create_shader(device, cs_uav_clear_2d_float_code, sizeof(cs_uav_clear_2d_float_code),
> + &state->float_shaders.image_2d);
> + create_shader(device, cs_uav_clear_2d_uint_code, sizeof(cs_uav_clear_2d_uint_code),
> + &state->uint_shaders.image_2d);
> + create_shader(device, cs_uav_clear_2d_array_float_code, sizeof(cs_uav_clear_2d_array_float_code),
> + &state->float_shaders.image_2d_array);
> + create_shader(device, cs_uav_clear_2d_array_uint_code, sizeof(cs_uav_clear_2d_array_uint_code),
> + &state->uint_shaders.image_2d_array);
> + create_shader(device, cs_uav_clear_3d_float_code, sizeof(cs_uav_clear_3d_float_code),
> + &state->float_shaders.image_3d);
> + create_shader(device, cs_uav_clear_3d_uint_code, sizeof(cs_uav_clear_3d_uint_code),
> + &state->uint_shaders.image_3d);
> +
> + wined3d_context_vk_create_bo(context_vk, sizeof(struct wined3d_uav_clear_constants_vk),
> + VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
> + &state->constants_bo);
> +}
> +
We never check the create_shader() return code.
> +void wined3d_device_vk_uav_clear_state_cleanup(struct wined3d_device_vk *device_vk)
> +{
> + struct wined3d_context_vk *context_vk = &device_vk->context_vk;
> + struct wined3d_uav_clear_state_vk *state = &device_vk->uav_clear_state;
> +
> + wined3d_context_vk_destroy_bo(context_vk, &state->constants_bo);
> +
> + if (state->float_shaders.buffer)
> + wined3d_shader_decref(state->float_shaders.buffer);
> + if (state->uint_shaders.buffer)
> + wined3d_shader_decref(state->uint_shaders.buffer);
> + if (state->float_shaders.image_1d)
> + wined3d_shader_decref(state->float_shaders.image_1d);
> + if (state->uint_shaders.image_1d)
> + wined3d_shader_decref(state->uint_shaders.image_1d);
> + if (state->float_shaders.image_1d_array)
> + wined3d_shader_decref(state->float_shaders.image_1d_array);
> + if (state->uint_shaders.image_1d_array)
> + wined3d_shader_decref(state->uint_shaders.image_1d_array);
> + if (state->float_shaders.image_2d)
> + wined3d_shader_decref(state->float_shaders.image_2d);
> + if (state->uint_shaders.image_2d)
> + wined3d_shader_decref(state->uint_shaders.image_2d);
> + if (state->float_shaders.image_2d_array)
> + wined3d_shader_decref(state->float_shaders.image_2d_array);
> + if (state->uint_shaders.image_2d_array)
> + wined3d_shader_decref(state->uint_shaders.image_2d_array);
> + if (state->float_shaders.image_3d)
> + wined3d_shader_decref(state->float_shaders.image_3d);
> + if (state->uint_shaders.image_3d)
> + wined3d_shader_decref(state->uint_shaders.image_3d);
> +}
> +
...but if we did, these shaders could never be NULL here.
> + constants.extent.width = resource->width;
> + constants.extent.height = resource->height;
> +
> + group_count = shader->u.cs.thread_group_size;
> +
> + if (resource->type != WINED3D_RTYPE_BUFFER)
> + {
> + constants.extent.width >>= view_desc->u.texture.level_idx;
> + constants.extent.height >>= view_desc->u.texture.level_idx;
> + group_count.z = (view_desc->u.texture.layer_count + group_count.z - 1) / group_count.z;
> + }
> +
I.e., wined3d_texture_get_level_width(), wined3d_texture_get_level_height().
> + constants_bo = &device_vk->uav_clear_state.constants_bo;
> + bo_address.buffer_object = (uintptr_t)constants_bo;
> + bo_address.addr = NULL;
> +
> + mapped_constants_bo = wined3d_context_map_bo_address(&context_vk->c, &bo_address,
> + sizeof(constants), WINED3D_MAP_WRITE | WINED3D_MAP_DISCARD);
> + memcpy(mapped_constants_bo, &constants, sizeof(constants));
> +
> + bo_range.offset = 0;
> + bo_range.size = sizeof(constants);
> + wined3d_context_unmap_bo_address(&context_vk->c, &bo_address, 1, &bo_range);
> +
I.e., wined3d_context_copy_bo_address().
> vk_barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
> vk_barrier.pNext = NULL;
> - vk_barrier.srcAccessMask = access_mask;
> - vk_barrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
> + vk_barrier.srcAccessMask = VK_ACCESS_HOST_WRITE_BIT;
> + vk_barrier.dstAccessMask = VK_ACCESS_UNIFORM_READ_BIT;
> vk_barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
> vk_barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
> - vk_barrier.buffer = buffer_vk->bo.vk_buffer;
> - vk_barrier.offset = buffer_vk->bo.buffer_offset + offset;
> - vk_barrier.size = size;
> - VK_CALL(vkCmdPipelineBarrier(vk_command_buffer, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
> - VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, NULL, 1, &vk_barrier, 0, NULL));
> + vk_barrier.buffer = constants_bo->vk_buffer;
> + vk_barrier.offset = constants_bo->buffer_offset;
> + vk_barrier.size = constants_bo->size;
> + VK_CALL(vkCmdPipelineBarrier(wined3d_context_vk_get_command_buffer(context_vk),
> + VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
> + 0, 0, NULL, 1, &vk_barrier, 0, NULL));
>
Do we need that barrier? Specifically, is this not already covered by
"Host Write Ordering Guarantees"?
More information about the wine-devel
mailing list