#include "util/format_rgb9e5.h"
#include "blorp_priv.h"
-#include "brw_defines.h"
+#include "compiler/brw_eu_defines.h"
#include "compiler/nir/nir_builder.h"
struct brw_blorp_const_color_prog_key
{
+ enum blorp_shader_type shader_type; /* Must be BLORP_SHADER_TYPE_CLEAR */
bool use_simd16_replicated_data;
bool pad[3];
};
struct blorp_params *params,
bool use_replicated_data)
{
- struct brw_blorp_const_color_prog_key blorp_key;
- memset(&blorp_key, 0, sizeof(blorp_key));
- blorp_key.use_simd16_replicated_data = use_replicated_data;
+ const struct brw_blorp_const_color_prog_key blorp_key = {
+ .shader_type = BLORP_SHADER_TYPE_CLEAR,
+ .use_simd16_replicated_data = use_replicated_data,
+ };
if (blorp->lookup_shader(blorp, &blorp_key, sizeof(blorp_key),
¶ms->wm_prog_kernel, ¶ms->wm_prog_data))
void *mem_ctx = ralloc_context(NULL);
nir_builder b;
- nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
- b.shader->info.name = ralloc_strdup(b.shader, "BLORP-clear");
+ nir_builder_init_simple_shader(&b, mem_ctx, MESA_SHADER_FRAGMENT, NULL);
+ b.shader->info->name = ralloc_strdup(b.shader, "BLORP-clear");
- nir_variable *v_color = nir_variable_create(b.shader, nir_var_shader_in,
- glsl_vec4_type(), "v_color");
- v_color->data.location = VARYING_SLOT_VAR0;
- v_color->data.interpolation = INTERP_MODE_FLAT;
+ nir_variable *v_color =
+ BLORP_CREATE_NIR_INPUT(b.shader, clear_color, glsl_vec4_type());
nir_variable *frag_color = nir_variable_create(b.shader, nir_var_shader_out,
glsl_vec4_type(),
struct brw_wm_prog_key wm_key;
brw_blorp_init_wm_prog_key(&wm_key);
- struct brw_blorp_prog_data prog_data;
+ struct brw_wm_prog_data prog_data;
unsigned program_size;
const unsigned *program =
- brw_blorp_compile_nir_shader(blorp, b.shader, &wm_key, use_replicated_data,
- &prog_data, &program_size);
+ blorp_compile_fs(blorp, mem_ctx, b.shader, &wm_key, use_replicated_data,
+ &prog_data, &program_size);
blorp->upload_shader(blorp, &blorp_key, sizeof(blorp_key),
program, program_size,
- &prog_data, sizeof(prog_data),
+ &prog_data.base, sizeof(prog_data),
¶ms->wm_prog_kernel, ¶ms->wm_prog_data);
ralloc_free(mem_ctx);
}
+struct layer_offset_vs_key {
+ enum blorp_shader_type shader_type;
+ unsigned num_inputs;
+};
+
+/* In the case of doing attachment clears, we are using a surface state that
+ * is handed to us so we can't set (and don't even know) the base array layer.
+ * In order to do a layered clear in this scenario, we need some way of adding
+ * the base array layer to the instance id. Unfortunately, our hardware has
+ * no real concept of "base instance", so we have to do it manually in a
+ * vertex shader.
+ */
+static void
+blorp_params_get_layer_offset_vs(struct blorp_context *blorp,
+ struct blorp_params *params)
+{
+ struct layer_offset_vs_key blorp_key = {
+ .shader_type = BLORP_SHADER_TYPE_LAYER_OFFSET_VS,
+ };
+
+ if (params->wm_prog_data)
+ blorp_key.num_inputs = params->wm_prog_data->num_varying_inputs;
+
+ if (blorp->lookup_shader(blorp, &blorp_key, sizeof(blorp_key),
+ ¶ms->vs_prog_kernel, ¶ms->vs_prog_data))
+ return;
+
+ void *mem_ctx = ralloc_context(NULL);
+
+ nir_builder b;
+ nir_builder_init_simple_shader(&b, mem_ctx, MESA_SHADER_VERTEX, NULL);
+ b.shader->info->name = ralloc_strdup(b.shader, "BLORP-layer-offset-vs");
+
+ const struct glsl_type *uvec4_type = glsl_vector_type(GLSL_TYPE_UINT, 4);
+
+ /* First we deal with the header which has instance and base instance */
+ nir_variable *a_header = nir_variable_create(b.shader, nir_var_shader_in,
+ uvec4_type, "header");
+ a_header->data.location = VERT_ATTRIB_GENERIC0;
+
+ nir_variable *v_layer = nir_variable_create(b.shader, nir_var_shader_out,
+ glsl_int_type(), "layer_id");
+ v_layer->data.location = VARYING_SLOT_LAYER;
+
+ /* Compute the layer id */
+ nir_ssa_def *header = nir_load_var(&b, a_header);
+ nir_ssa_def *base_layer = nir_channel(&b, header, 0);
+ nir_ssa_def *instance = nir_channel(&b, header, 1);
+ nir_store_var(&b, v_layer, nir_iadd(&b, instance, base_layer), 0x1);
+
+ /* Then we copy the vertex from the next slot to VARYING_SLOT_POS */
+ nir_variable *a_vertex = nir_variable_create(b.shader, nir_var_shader_in,
+ glsl_vec4_type(), "a_vertex");
+ a_vertex->data.location = VERT_ATTRIB_GENERIC1;
+
+ nir_variable *v_pos = nir_variable_create(b.shader, nir_var_shader_out,
+ glsl_vec4_type(), "v_pos");
+ v_pos->data.location = VARYING_SLOT_POS;
+
+ nir_copy_var(&b, v_pos, a_vertex);
+
+ /* Then we copy everything else */
+ for (unsigned i = 0; i < blorp_key.num_inputs; i++) {
+ nir_variable *a_in = nir_variable_create(b.shader, nir_var_shader_in,
+ uvec4_type, "input");
+ a_in->data.location = VERT_ATTRIB_GENERIC2 + i;
+
+ nir_variable *v_out = nir_variable_create(b.shader, nir_var_shader_out,
+ uvec4_type, "output");
+ v_out->data.location = VARYING_SLOT_VAR0 + i;
+
+ nir_copy_var(&b, v_out, a_in);
+ }
+
+ struct brw_vs_prog_data vs_prog_data;
+ memset(&vs_prog_data, 0, sizeof(vs_prog_data));
+
+ unsigned program_size;
+ const unsigned *program =
+ blorp_compile_vs(blorp, mem_ctx, b.shader, &vs_prog_data, &program_size);
+
+ blorp->upload_shader(blorp, &blorp_key, sizeof(blorp_key),
+ program, program_size,
+ &vs_prog_data.base.base, sizeof(vs_prog_data),
+ ¶ms->vs_prog_kernel, ¶ms->vs_prog_data);
+
+ ralloc_free(mem_ctx);
+}
+
/* The x0, y0, x1, and y1 parameters must already be populated with the render
* area of the framebuffer to be cleared.
*/
params.x1 = x1;
params.y1 = y1;
- memset(¶ms.wm_inputs, 0xff, 4*sizeof(float));
+ memset(¶ms.wm_inputs.clear_color, 0xff, 4*sizeof(float));
params.fast_clear_op = BLORP_FAST_CLEAR_OP_CLEAR;
get_fast_clear_rect(batch->blorp->isl_dev, surf->aux_surf,
brw_blorp_surface_info_init(batch->blorp, ¶ms.dst, surf, level,
start_layer, format, true);
+ params.num_samples = params.dst.surf.samples;
batch->blorp->exec(batch, ¶ms);
}
+static union isl_color_value
+swizzle_color_value(union isl_color_value src, struct isl_swizzle swizzle)
+{
+ union isl_color_value dst = { .u32 = { 0, } };
+
+ /* We assign colors in ABGR order so that the first one will be taken in
+ * RGBA precedence order. According to the PRM docs for shader channel
+ * select, this matches Haswell hardware behavior.
+ */
+ if ((unsigned)(swizzle.a - ISL_CHANNEL_SELECT_RED) < 4)
+ dst.u32[swizzle.a - ISL_CHANNEL_SELECT_RED] = src.u32[3];
+ if ((unsigned)(swizzle.b - ISL_CHANNEL_SELECT_RED) < 4)
+ dst.u32[swizzle.b - ISL_CHANNEL_SELECT_RED] = src.u32[2];
+ if ((unsigned)(swizzle.g - ISL_CHANNEL_SELECT_RED) < 4)
+ dst.u32[swizzle.g - ISL_CHANNEL_SELECT_RED] = src.u32[1];
+ if ((unsigned)(swizzle.r - ISL_CHANNEL_SELECT_RED) < 4)
+ dst.u32[swizzle.r - ISL_CHANNEL_SELECT_RED] = src.u32[0];
+
+ return dst;
+}
void
blorp_clear(struct blorp_batch *batch,
const struct blorp_surf *surf,
+ enum isl_format format, struct isl_swizzle swizzle,
uint32_t level, uint32_t start_layer, uint32_t num_layers,
uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
- enum isl_format format, union isl_color_value clear_color,
- bool color_write_disable[4])
+ union isl_color_value clear_color,
+ const bool color_write_disable[4])
{
struct blorp_params params;
blorp_params_init(¶ms);
params.x1 = x1;
params.y1 = y1;
+ /* Manually apply the clear destination swizzle. This way swizzled clears
+ * will work for swizzles which we can't normally use for rendering and it
+ * also ensures that they work on pre-Haswell hardware which can't swizlle
+ * at all.
+ */
+ clear_color = swizzle_color_value(clear_color, swizzle);
+ swizzle = ISL_SWIZZLE_IDENTITY;
+
if (format == ISL_FORMAT_R9G9B9E5_SHAREDEXP) {
clear_color.u32[0] = float3_to_rgb9e5(clear_color.f32);
format = ISL_FORMAT_R32_UINT;
+ } else if (format == ISL_FORMAT_A4B4G4R4_UNORM) {
+ /* Broadwell and earlier cannot render to this format so we need to work
+ * around it by swapping the colors around and using B4G4R4A4 instead.
+ */
+ const struct isl_swizzle ARGB = ISL_SWIZZLE(ALPHA, RED, GREEN, BLUE);
+ clear_color = swizzle_color_value(clear_color, ARGB);
+ format = ISL_FORMAT_B4G4R4A4_UNORM;
}
- memcpy(¶ms.wm_inputs, clear_color.f32, sizeof(float) * 4);
+ memcpy(¶ms.wm_inputs.clear_color, clear_color.f32, sizeof(float) * 4);
bool use_simd16_replicated_data = true;
/* Constant color writes ignore everyting in blend and color calculator
* state. This is not documented.
*/
- for (unsigned i = 0; i < 4; i++) {
- params.color_write_disable[i] = color_write_disable[i];
- if (color_write_disable[i])
- use_simd16_replicated_data = false;
+ if (color_write_disable) {
+ for (unsigned i = 0; i < 4; i++) {
+ params.color_write_disable[i] = color_write_disable[i];
+ if (color_write_disable[i])
+ use_simd16_replicated_data = false;
+ }
}
blorp_params_get_clear_kernel(batch->blorp, ¶ms,
while (num_layers > 0) {
brw_blorp_surface_info_init(batch->blorp, ¶ms.dst, surf, level,
start_layer, format, true);
+ params.dst.view.swizzle = swizzle;
+
+ params.num_samples = params.dst.surf.samples;
/* We may be restricted on the number of layers we can bind at any one
* time. In particular, Sandy Bridge has a maximum number of layers of
}
}
+void
+blorp_clear_depth_stencil(struct blorp_batch *batch,
+ const struct blorp_surf *depth,
+ const struct blorp_surf *stencil,
+ uint32_t level, uint32_t start_layer,
+ uint32_t num_layers,
+ uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
+ bool clear_depth, float depth_value,
+ uint8_t stencil_mask, uint8_t stencil_value)
+{
+ struct blorp_params params;
+ blorp_params_init(¶ms);
+
+ params.x0 = x0;
+ params.y0 = y0;
+ params.x1 = x1;
+ params.y1 = y1;
+
+ while (num_layers > 0) {
+ params.num_layers = num_layers;
+
+ if (stencil_mask) {
+ brw_blorp_surface_info_init(batch->blorp, ¶ms.stencil, stencil,
+ level, start_layer,
+ ISL_FORMAT_UNSUPPORTED, true);
+ params.stencil_mask = stencil_mask;
+ params.stencil_ref = stencil_value;
+
+ params.dst.surf.samples = params.stencil.surf.samples;
+ params.dst.surf.logical_level0_px =
+ params.stencil.surf.logical_level0_px;
+ params.dst.view = params.depth.view;
+
+ params.num_samples = params.stencil.surf.samples;
+
+ /* We may be restricted on the number of layers we can bind at any
+ * one time. In particular, Sandy Bridge has a maximum number of
+ * layers of 512 but a maximum 3D texture size is much larger.
+ */
+ if (params.stencil.view.array_len < params.num_layers)
+ params.num_layers = params.stencil.view.array_len;
+ }
+
+ if (clear_depth) {
+ brw_blorp_surface_info_init(batch->blorp, ¶ms.depth, depth,
+ level, start_layer,
+ ISL_FORMAT_UNSUPPORTED, true);
+ params.z = depth_value;
+ params.depth_format =
+ isl_format_get_depth_format(depth->surf->format, false);
+
+ params.dst.surf.samples = params.depth.surf.samples;
+ params.dst.surf.logical_level0_px =
+ params.depth.surf.logical_level0_px;
+ params.dst.view = params.depth.view;
+
+ params.num_samples = params.depth.surf.samples;
+
+ /* We may be restricted on the number of layers we can bind at any
+ * one time. In particular, Sandy Bridge has a maximum number of
+ * layers of 512 but a maximum 3D texture size is much larger.
+ */
+ if (params.depth.view.array_len < params.num_layers)
+ params.num_layers = params.depth.view.array_len;
+ }
+
+ batch->blorp->exec(batch, ¶ms);
+
+ start_layer += params.num_layers;
+ num_layers -= params.num_layers;
+ }
+}
+
+bool
+blorp_can_hiz_clear_depth(uint8_t gen, enum isl_format format,
+ uint32_t num_samples,
+ uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1)
+{
+ /* This function currently doesn't support any gen prior to gen8 */
+ assert(gen >= 8);
+
+ if (gen == 8 && format == ISL_FORMAT_R16_UNORM) {
+ /* Apply the D16 alignment restrictions. On BDW, HiZ has an 8x4 sample
+ * block with the following property: as the number of samples increases,
+ * the number of pixels representable by this block decreases by a factor
+ * of the sample dimensions. Sample dimensions scale following the MSAA
+ * interleaved pattern.
+ *
+ * Sample|Sample|Pixel
+ * Count |Dim |Dim
+ * ===================
+ * 1 | 1x1 | 8x4
+ * 2 | 2x1 | 4x4
+ * 4 | 2x2 | 4x2
+ * 8 | 4x2 | 2x2
+ * 16 | 4x4 | 2x1
+ *
+ * Table: Pixel Dimensions in a HiZ Sample Block Pre-SKL
+ */
+ const struct isl_extent2d sa_block_dim =
+ isl_get_interleaved_msaa_px_size_sa(num_samples);
+ const uint8_t align_px_w = 8 / sa_block_dim.w;
+ const uint8_t align_px_h = 4 / sa_block_dim.h;
+
+ /* Fast depth clears clear an entire sample block at a time. As a result,
+ * the rectangle must be aligned to the dimensions of the encompassing
+ * pixel block for a successful operation.
+ *
+ * Fast clears can still work if the upper-left corner is aligned and the
+ * bottom-rigtht corner touches the edge of a depth buffer whose extent
+ * is unaligned. This is because each miplevel in the depth buffer is
+ * padded by the Pixel Dim (similar to a standard compressed texture).
+ * In this case, the clear rectangle could be padded by to match the full
+ * depth buffer extent but to support multiple clearing techniques, we
+ * chose to be unaware of the depth buffer's extent and thus don't handle
+ * this case.
+ */
+ if (x0 % align_px_w || y0 % align_px_h ||
+ x1 % align_px_w || y1 % align_px_h)
+ return false;
+ }
+ return true;
+}
+
+/* Given a depth stencil attachment, this function performs a fast depth clear
+ * on a depth portion and a regular clear on the stencil portion. When
+ * performing a fast depth clear on the depth portion, the HiZ buffer is simply
+ * tagged as cleared so the depth clear value is not actually needed.
+ */
+void
+blorp_gen8_hiz_clear_attachments(struct blorp_batch *batch,
+ uint32_t num_samples,
+ uint32_t x0, uint32_t y0,
+ uint32_t x1, uint32_t y1,
+ bool clear_depth, bool clear_stencil,
+ uint8_t stencil_value)
+{
+ assert(batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL);
+
+ struct blorp_params params;
+ blorp_params_init(¶ms);
+ params.num_layers = 1;
+ params.hiz_op = BLORP_HIZ_OP_DEPTH_CLEAR;
+ params.x0 = x0;
+ params.y0 = y0;
+ params.x1 = x1;
+ params.y1 = y1;
+ params.num_samples = num_samples;
+ params.depth.enabled = clear_depth;
+ params.stencil.enabled = clear_stencil;
+ params.stencil_ref = stencil_value;
+ batch->blorp->exec(batch, ¶ms);
+}
+
+/** Clear active color/depth/stencili attachments
+ *
+ * This function performs a clear operation on the currently bound
+ * color/depth/stencil attachments. It is assumed that any information passed
+ * in here is valid, consistent, and in-bounds relative to the currently
+ * attached depth/stencil. The binding_table_offset parameter is the 32-bit
+ * offset relative to surface state base address where pre-baked binding table
+ * that we are to use lives. If clear_color is false, binding_table_offset
+ * must point to a binding table with one entry which is a valid null surface
+ * that matches the currently bound depth and stencil.
+ */
+void
+blorp_clear_attachments(struct blorp_batch *batch,
+ uint32_t binding_table_offset,
+ enum isl_format depth_format,
+ uint32_t num_samples,
+ uint32_t start_layer, uint32_t num_layers,
+ uint32_t x0, uint32_t y0, uint32_t x1, uint32_t y1,
+ bool clear_color, union isl_color_value color_value,
+ bool clear_depth, float depth_value,
+ uint8_t stencil_mask, uint8_t stencil_value)
+{
+ struct blorp_params params;
+ blorp_params_init(¶ms);
+
+ assert(batch->flags & BLORP_BATCH_NO_EMIT_DEPTH_STENCIL);
+
+ params.x0 = x0;
+ params.y0 = y0;
+ params.x1 = x1;
+ params.y1 = y1;
+
+ params.use_pre_baked_binding_table = true;
+ params.pre_baked_binding_table_offset = binding_table_offset;
+
+ params.num_layers = num_layers;
+ params.num_samples = num_samples;
+
+ if (clear_color) {
+ params.dst.enabled = true;
+
+ memcpy(¶ms.wm_inputs.clear_color, color_value.f32, sizeof(float) * 4);
+
+ /* Unfortunately, without knowing whether or not our destination surface
+ * is tiled or not, we have to assume it may be linear. This means no
+ * SIMD16_REPDATA for us. :-(
+ */
+ blorp_params_get_clear_kernel(batch->blorp, ¶ms, false);
+ }
+
+ if (clear_depth) {
+ params.depth.enabled = true;
+
+ params.z = depth_value;
+ params.depth_format = isl_format_get_depth_format(depth_format, false);
+ }
+
+ if (stencil_mask) {
+ params.stencil.enabled = true;
+
+ params.stencil_mask = stencil_mask;
+ params.stencil_ref = stencil_value;
+ }
+
+ blorp_params_get_layer_offset_vs(batch->blorp, ¶ms);
+ params.vs_inputs.base_layer = start_layer;
+
+ batch->blorp->exec(batch, ¶ms);
+}
+
void
blorp_ccs_resolve(struct blorp_batch *batch,
- struct blorp_surf *surf, enum isl_format format)
+ struct blorp_surf *surf, uint32_t level, uint32_t layer,
+ enum isl_format format,
+ enum blorp_fast_clear_op resolve_op)
{
struct blorp_params params;
blorp_params_init(¶ms);
+ /* Layered and mipmapped fast clear is only available from Gen8 onwards. */
+ assert(ISL_DEV_GEN(batch->blorp->isl_dev) >= 8 ||
+ (level == 0 && layer == 0));
+
brw_blorp_surface_info_init(batch->blorp, ¶ms.dst, surf,
- 0 /* level */, 0 /* layer */, format, true);
+ level, layer, format, true);
/* From the Ivy Bridge PRM, Vol2 Part1 11.9 "Render Target Resolve":
*
y_scaledown = aux_fmtl->bh / 2;
}
params.x0 = params.y0 = 0;
- params.x1 = params.dst.aux_surf.logical_level0_px.width;
- params.y1 = params.dst.aux_surf.logical_level0_px.height;
+ params.x1 = minify(params.dst.aux_surf.logical_level0_px.width, level);
+ params.y1 = minify(params.dst.aux_surf.logical_level0_px.height, level);
params.x1 = ALIGN(params.x1, x_scaledown) / x_scaledown;
params.y1 = ALIGN(params.y1, y_scaledown) / y_scaledown;
if (batch->blorp->isl_dev->info->gen >= 9) {
- if (params.dst.aux_usage == ISL_AUX_USAGE_CCS_E)
- params.fast_clear_op = BLORP_FAST_CLEAR_OP_RESOLVE_FULL;
- else
- params.fast_clear_op = BLORP_FAST_CLEAR_OP_RESOLVE_PARTIAL;
+ assert(resolve_op == BLORP_FAST_CLEAR_OP_RESOLVE_FULL ||
+ resolve_op == BLORP_FAST_CLEAR_OP_RESOLVE_PARTIAL);
} else {
/* Broadwell and earlier do not have a partial resolve */
- params.fast_clear_op = BLORP_FAST_CLEAR_OP_RESOLVE_FULL;
+ assert(resolve_op == BLORP_FAST_CLEAR_OP_RESOLVE_FULL);
}
+ params.fast_clear_op = resolve_op;
/* Note: there is no need to initialize push constants because it doesn't
* matter what data gets dispatched to the render target. However, we must