#include "compiler/nir/nir_builder.h"
#include "blorp_priv.h"
-#include "brw_meta_util.h"
+
+/* header-only include needed for _mesa_unorm_to_float and friends. */
+#include "mesa/main/format_utils.h"
#define FILE_DEBUG_FLAG DEBUG_BLORP
+static const bool split_blorp_blit_debug = false;
+
/**
* Enum to specify the order of arguments in a sampler message
*/
assert(!key->use_kill || !(key->blend && key->blit_scaled));
#define LOAD_INPUT(name, type)\
- v->v_##name = nir_variable_create(b->shader, nir_var_shader_in, \
- type, #name); \
- v->v_##name->data.interpolation = INTERP_MODE_FLAT; \
- v->v_##name->data.location = VARYING_SLOT_VAR0 + \
- offsetof(struct brw_blorp_wm_inputs, name) / (4 * sizeof(float)); \
- v->v_##name->data.location_frac = \
- (offsetof(struct brw_blorp_wm_inputs, name) / sizeof(float)) % 4;
+ v->v_##name = BLORP_CREATE_NIR_INPUT(b->shader, name, type);
LOAD_INPUT(discard_rect, glsl_vec4_type())
LOAD_INPUT(rect_grid, glsl_vec4_type())
const struct brw_blorp_blit_prog_key *key,
struct brw_blorp_blit_vars *v)
{
- nir_ssa_def *coord = nir_f2i(b, nir_load_var(b, v->frag_coord));
+ nir_ssa_def *coord = nir_f2i32(b, nir_load_var(b, v->frag_coord));
/* Account for destination surface intratile offset
*
nir_ssa_def *sample_off = nir_imm_vec2(b, sample_off_x, sample_off_y);
nir_ssa_def *sample_coords = nir_fadd(b, pos_xy, sample_off);
- nir_ssa_def *sample_coords_int = nir_f2i(b, sample_coords);
+ nir_ssa_def *sample_coords_int = nir_f2i32(b, sample_coords);
/* The MCS value we fetch has to match up with the pixel that we're
* sampling from. Since we sample from different pixels in each
nir_ssa_def *sample =
nir_fdot2(b, frac, nir_imm_vec2(b, key->x_scale,
key->x_scale * key->y_scale));
- sample = nir_f2i(b, sample);
+ sample = nir_f2i32(b, sample);
if (tex_samples == 8) {
sample = nir_iand(b, nir_ishr(b, nir_imm_int(b, 0x64210573),
frac_y);
}
+/** Perform a color bit-cast operation
+ *
+ * For copy operations involving CCS, we may need to use different formats for
+ * the source and destination surfaces. The two formats must both be UINT
+ * formats and must have the same size but may have different bit layouts.
+ * For instance, we may be copying from R8G8B8A8_UINT to R32_UINT or R32_UINT
+ * to R16G16_UINT. This function generates code to shuffle bits around to get
+ * us from one to the other.
+ */
+static nir_ssa_def *
+bit_cast_color(struct nir_builder *b, nir_ssa_def *color,
+ const struct brw_blorp_blit_prog_key *key)
+{
+ assert(key->texture_data_type == nir_type_uint);
+
+ if (key->dst_bpc > key->src_bpc) {
+ nir_ssa_def *u = nir_ssa_undef(b, 1, 32);
+ nir_ssa_def *dst_chan[2] = { u, u };
+ unsigned shift = 0;
+ unsigned dst_idx = 0;
+ for (unsigned i = 0; i < 4; i++) {
+ nir_ssa_def *shifted = nir_ishl(b, nir_channel(b, color, i),
+ nir_imm_int(b, shift));
+ if (shift == 0) {
+ dst_chan[dst_idx] = shifted;
+ } else {
+ dst_chan[dst_idx] = nir_ior(b, dst_chan[dst_idx], shifted);
+ }
+
+ shift += key->src_bpc;
+ if (shift >= key->dst_bpc) {
+ dst_idx++;
+ shift = 0;
+ }
+ }
+
+ return nir_vec4(b, dst_chan[0], dst_chan[1], u, u);
+ } else {
+ assert(key->dst_bpc < key->src_bpc);
+
+ nir_ssa_def *mask = nir_imm_int(b, ~0u >> (32 - key->dst_bpc));
+
+ nir_ssa_def *dst_chan[4];
+ unsigned src_idx = 0;
+ unsigned shift = 0;
+ for (unsigned i = 0; i < 4; i++) {
+ dst_chan[i] = nir_iand(b, nir_ushr(b, nir_channel(b, color, src_idx),
+ nir_imm_int(b, shift)),
+ mask);
+ shift += key->dst_bpc;
+ if (shift >= key->src_bpc) {
+ src_idx++;
+ shift = 0;
+ }
+ }
+
+ return nir_vec4(b, dst_chan[0], dst_chan[1], dst_chan[2], dst_chan[3]);
+ }
+}
+
/**
* Generator for WM programs used in BLORP blits.
*
* of samples).
*/
static nir_shader *
-brw_blorp_build_nir_shader(struct blorp_context *blorp,
+brw_blorp_build_nir_shader(struct blorp_context *blorp, void *mem_ctx,
const struct brw_blorp_blit_prog_key *key)
{
const struct gen_device_info *devinfo = blorp->isl_dev->info;
(key->dst_samples <= 1));
nir_builder b;
- nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
+ nir_builder_init_simple_shader(&b, mem_ctx, MESA_SHADER_FRAGMENT, NULL);
struct brw_blorp_blit_vars v;
brw_blorp_blit_vars_init(&b, &v, key);
blorp_nir_discard_if_outside_rect(&b, dst_pos, &v);
}
- src_pos = blorp_blit_apply_transform(&b, nir_i2f(&b, dst_pos), &v);
+ src_pos = blorp_blit_apply_transform(&b, nir_i2f32(&b, dst_pos), &v);
if (dst_pos->num_components == 3) {
/* The sample coordinate is an integer that we want left alone but
* blorp_blit_apply_transform() blindly applies the transform to all
/* Resolves (effecively) use texelFetch, so we need integers and we
* don't care about the sample index if we got one.
*/
- src_pos = nir_f2i(&b, nir_channels(&b, src_pos, 0x3));
+ src_pos = nir_f2i32(&b, nir_channels(&b, src_pos, 0x3));
if (devinfo->gen == 6) {
/* Because gen6 only supports 4x interleved MSAA, we can do all the
*/
src_pos = nir_ishl(&b, src_pos, nir_imm_int(&b, 1));
src_pos = nir_iadd(&b, src_pos, nir_imm_int(&b, 1));
- src_pos = nir_i2f(&b, src_pos);
+ src_pos = nir_i2f32(&b, src_pos);
color = blorp_nir_tex(&b, &v, src_pos, key->texture_data_type);
} else {
/* Gen7+ hardware doesn't automaticaly blend. */
} else {
/* We're going to use texelFetch, so we need integers */
if (src_pos->num_components == 2) {
- src_pos = nir_f2i(&b, src_pos);
+ src_pos = nir_f2i32(&b, src_pos);
} else {
assert(src_pos->num_components == 3);
- src_pos = nir_vec3(&b, nir_channel(&b, nir_f2i(&b, src_pos), 0),
- nir_channel(&b, nir_f2i(&b, src_pos), 1),
+ src_pos = nir_vec3(&b, nir_channel(&b, nir_f2i32(&b, src_pos), 0),
+ nir_channel(&b, nir_f2i32(&b, src_pos), 1),
nir_channel(&b, src_pos, 2));
}
}
}
+ if (key->dst_bpc != key->src_bpc)
+ color = bit_cast_color(&b, color, key);
+
+ if (key->dst_rgb) {
+ /* The destination image is bound as a red texture three times as wide
+ * as the actual image. Our shader is effectively running one color
+ * component at a time. We need to pick off the appropriate component
+ * from the source color and write that to destination red.
+ */
+ assert(dst_pos->num_components == 2);
+ nir_ssa_def *comp =
+ nir_umod(&b, nir_channel(&b, dst_pos, 0), nir_imm_int(&b, 3));
+
+ nir_ssa_def *color_component =
+ nir_bcsel(&b, nir_ieq(&b, comp, nir_imm_int(&b, 0)),
+ nir_channel(&b, color, 0),
+ nir_bcsel(&b, nir_ieq(&b, comp, nir_imm_int(&b, 1)),
+ nir_channel(&b, color, 1),
+ nir_channel(&b, color, 2)));
+
+ nir_ssa_def *u = nir_ssa_undef(&b, 1, 32);
+ color = nir_vec4(&b, color_component, u, u, u);
+ }
+
nir_store_var(&b, v.color_out, color, 0xf);
return b.shader;
}
-static void
+static bool
brw_blorp_get_blit_kernel(struct blorp_context *blorp,
struct blorp_params *params,
const struct brw_blorp_blit_prog_key *prog_key)
{
if (blorp->lookup_shader(blorp, prog_key, sizeof(*prog_key),
¶ms->wm_prog_kernel, ¶ms->wm_prog_data))
- return;
+ return true;
+
+ void *mem_ctx = ralloc_context(NULL);
const unsigned *program;
unsigned program_size;
- struct brw_blorp_prog_data prog_data;
+ struct brw_wm_prog_data prog_data;
+
+ nir_shader *nir = brw_blorp_build_nir_shader(blorp, mem_ctx, prog_key);
+ nir->info->name = ralloc_strdup(nir, "BLORP-blit");
- /* Try and compile with NIR first. If that fails, fall back to the old
- * method of building shaders manually.
- */
- nir_shader *nir = brw_blorp_build_nir_shader(blorp, prog_key);
struct brw_wm_prog_key wm_key;
brw_blorp_init_wm_prog_key(&wm_key);
wm_key.tex.compressed_multisample_layout_mask =
wm_key.tex.msaa_16 = prog_key->tex_samples == 16;
wm_key.multisample_fbo = prog_key->rt_samples > 1;
- program = brw_blorp_compile_nir_shader(blorp, nir, &wm_key, false,
- &prog_data, &program_size);
+ program = blorp_compile_fs(blorp, mem_ctx, nir, &wm_key, false,
+ &prog_data, &program_size);
+
+ bool result =
+ blorp->upload_shader(blorp, prog_key, sizeof(*prog_key),
+ program, program_size,
+ &prog_data.base, sizeof(prog_data),
+ ¶ms->wm_prog_kernel, ¶ms->wm_prog_data);
- blorp->upload_shader(blorp, prog_key, sizeof(*prog_key),
- program, program_size,
- &prog_data, sizeof(prog_data),
- ¶ms->wm_prog_kernel, ¶ms->wm_prog_data);
+ ralloc_free(mem_ctx);
+ return result;
}
static void
GLfloat dst0, GLfloat dst1,
bool mirror)
{
- float scale = (src1 - src0) / (dst1 - dst0);
+ double scale = (double)(src1 - src0) / (double)(dst1 - dst0);
if (!mirror) {
/* When not mirroring a coordinate (say, X), we need:
* src_x - src_x0 = (dst_x - dst_x0 + 0.5) * scale
* so 0.5 provides the necessary correction.
*/
xform->multiplier = scale;
- xform->offset = src0 + (-dst0 + 0.5f) * scale;
+ xform->offset = src0 + (-(double)dst0 + 0.5) * scale;
} else {
/* When mirroring X we need:
* src_x - src_x0 = dst_x1 - dst_x - 0.5
* src_x = src_x0 + (dst_x1 -dst_x - 0.5) * scale
*/
xform->multiplier = -scale;
- xform->offset = src0 + (dst1 - 0.5f) * scale;
+ xform->offset = src0 + ((double)dst1 - 0.5) * scale;
}
}
surf_convert_to_single_slice(const struct isl_device *isl_dev,
struct brw_blorp_surface_info *info)
{
+ bool ok UNUSED;
+
/* Just bail if we have nothing to do. */
if (info->surf.dim == ISL_SURF_DIM_2D &&
info->view.base_level == 0 && info->view.base_array_layer == 0 &&
uint32_t tile_x_px, tile_y_px;
surf_get_intratile_offset_px(info, &tile_x_px, &tile_y_px);
- /* TODO: Once this file gets converted to C, we shouls just use designated
- * initializers.
- */
- struct isl_surf_init_info init_info = { 0, };
-
- init_info.dim = ISL_SURF_DIM_2D;
- init_info.format = info->surf.format;
- init_info.width = slice_width_px + tile_x_px;
- init_info.height = slice_height_px + tile_y_px;
- init_info.depth = 1;
- init_info.levels = 1;
- init_info.array_len = 1;
- init_info.samples = info->surf.samples;
- init_info.min_pitch = info->surf.row_pitch;
- init_info.usage = info->surf.usage;
- init_info.tiling_flags = 1 << info->surf.tiling;
-
- isl_surf_init_s(isl_dev, &info->surf, &init_info);
- assert(info->surf.row_pitch == init_info.min_pitch);
+ struct isl_surf_init_info init_info = {
+ .dim = ISL_SURF_DIM_2D,
+ .format = info->surf.format,
+ .width = slice_width_px + tile_x_px,
+ .height = slice_height_px + tile_y_px,
+ .depth = 1,
+ .levels = 1,
+ .array_len = 1,
+ .samples = info->surf.samples,
+ .row_pitch = info->surf.row_pitch,
+ .usage = info->surf.usage,
+ .tiling_flags = 1 << info->surf.tiling,
+ };
+
+ ok = isl_surf_init_s(isl_dev, &info->surf, &init_info);
+ assert(ok);
/* The view is also different now. */
info->view.base_level = 0;
info->tile_y_sa /= 2;
}
+static bool
+can_shrink_surface(const struct brw_blorp_surface_info *surf)
+{
+ /* The current code doesn't support offsets into the aux buffers. This
+ * should be possible, but we need to make sure the offset is page
+ * aligned for both the surface and the aux buffer surface. Generally
+ * this mean using the page aligned offset for the aux buffer.
+ *
+ * Currently the cases where we must split the blit are limited to cases
+ * where we don't have a aux buffer.
+ */
+ if (surf->aux_addr.buffer != NULL)
+ return false;
+
+ /* We can't support splitting the blit for gen <= 7, because the qpitch
+ * size is calculated by the hardware based on the surface height for
+ * gen <= 7. In gen >= 8, the qpitch is controlled by the driver.
+ */
+ if (surf->surf.msaa_layout == ISL_MSAA_LAYOUT_ARRAY)
+ return false;
+
+ return true;
+}
+
+static bool
+can_shrink_surfaces(const struct blorp_params *params)
+{
+ return
+ can_shrink_surface(¶ms->src) &&
+ can_shrink_surface(¶ms->dst);
+}
+
+static unsigned
+get_max_surface_size(const struct gen_device_info *devinfo,
+ const struct blorp_params *params)
+{
+ const unsigned max = devinfo->gen >= 7 ? 16384 : 8192;
+ if (split_blorp_blit_debug && can_shrink_surfaces(params))
+ return max >> 4; /* A smaller restriction when debug is enabled */
+ else
+ return max;
+}
+
+struct blt_axis {
+ double src0, src1, dst0, dst1;
+ bool mirror;
+};
+
+struct blt_coords {
+ struct blt_axis x, y;
+};
+
static void
-do_blorp_blit(struct blorp_batch *batch,
- struct blorp_params *params,
- struct brw_blorp_blit_prog_key *wm_prog_key,
- float src_x0, float src_y0,
- float src_x1, float src_y1,
- float dst_x0, float dst_y0,
- float dst_x1, float dst_y1,
- bool mirror_x, bool mirror_y)
+surf_fake_rgb_with_red(const struct isl_device *isl_dev,
+ struct brw_blorp_surface_info *info,
+ uint32_t *x, uint32_t *width)
+{
+ surf_convert_to_single_slice(isl_dev, info);
+
+ info->surf.logical_level0_px.width *= 3;
+ info->surf.phys_level0_sa.width *= 3;
+ *x *= 3;
+ *width *= 3;
+
+ enum isl_format red_format;
+ switch (info->view.format) {
+ case ISL_FORMAT_R8G8B8_UNORM:
+ red_format = ISL_FORMAT_R8_UNORM;
+ break;
+ case ISL_FORMAT_R8G8B8_UINT:
+ red_format = ISL_FORMAT_R8_UINT;
+ break;
+ case ISL_FORMAT_R16G16B16_UNORM:
+ red_format = ISL_FORMAT_R16_UNORM;
+ break;
+ case ISL_FORMAT_R16G16B16_UINT:
+ red_format = ISL_FORMAT_R16_UINT;
+ break;
+ case ISL_FORMAT_R32G32B32_UINT:
+ red_format = ISL_FORMAT_R32_UINT;
+ break;
+ default:
+ unreachable("Invalid RGB copy destination format");
+ }
+ assert(isl_format_get_layout(red_format)->channels.r.type ==
+ isl_format_get_layout(info->view.format)->channels.r.type);
+ assert(isl_format_get_layout(red_format)->channels.r.bits ==
+ isl_format_get_layout(info->view.format)->channels.r.bits);
+
+ info->surf.format = info->view.format = red_format;
+}
+
+static void
+fake_dest_rgb_with_red(const struct isl_device *dev,
+ struct blorp_params *params,
+ struct brw_blorp_blit_prog_key *wm_prog_key,
+ struct blt_coords *coords)
+{
+ /* Handle RGB destinations for blorp_copy */
+ const struct isl_format_layout *dst_fmtl =
+ isl_format_get_layout(params->dst.surf.format);
+
+ if (dst_fmtl->bpb % 3 == 0) {
+ uint32_t dst_x = coords->x.dst0;
+ uint32_t dst_width = coords->x.dst1 - dst_x;
+ surf_fake_rgb_with_red(dev, ¶ms->dst,
+ &dst_x, &dst_width);
+ coords->x.dst0 = dst_x;
+ coords->x.dst1 = dst_x + dst_width;
+ wm_prog_key->dst_rgb = true;
+ wm_prog_key->need_dst_offset = true;
+ }
+}
+
+enum blit_shrink_status {
+ BLIT_NO_SHRINK = 0,
+ BLIT_WIDTH_SHRINK = 1,
+ BLIT_HEIGHT_SHRINK = 2,
+};
+
+/* Try to blit. If the surface parameters exceed the size allowed by hardware,
+ * then enum blit_shrink_status will be returned. If BLIT_NO_SHRINK is
+ * returned, then the blit was successful.
+ */
+static enum blit_shrink_status
+try_blorp_blit(struct blorp_batch *batch,
+ struct blorp_params *params,
+ struct brw_blorp_blit_prog_key *wm_prog_key,
+ struct blt_coords *coords)
{
const struct gen_device_info *devinfo = batch->blorp->isl_dev->info;
+ fake_dest_rgb_with_red(batch->blorp->isl_dev, params, wm_prog_key, coords);
+
if (isl_format_has_sint_channel(params->src.view.format)) {
wm_prog_key->texture_data_type = nir_type_int;
} else if (isl_format_has_uint_channel(params->src.view.format)) {
/* Round floating point values to nearest integer to avoid "off by one texel"
* kind of errors when blitting.
*/
- params->x0 = params->wm_inputs.discard_rect.x0 = roundf(dst_x0);
- params->y0 = params->wm_inputs.discard_rect.y0 = roundf(dst_y0);
- params->x1 = params->wm_inputs.discard_rect.x1 = roundf(dst_x1);
- params->y1 = params->wm_inputs.discard_rect.y1 = roundf(dst_y1);
+ params->x0 = params->wm_inputs.discard_rect.x0 = round(coords->x.dst0);
+ params->y0 = params->wm_inputs.discard_rect.y0 = round(coords->y.dst0);
+ params->x1 = params->wm_inputs.discard_rect.x1 = round(coords->x.dst1);
+ params->y1 = params->wm_inputs.discard_rect.y1 = round(coords->y.dst1);
brw_blorp_setup_coord_transform(¶ms->wm_inputs.coord_transform[0],
- src_x0, src_x1, dst_x0, dst_x1, mirror_x);
+ coords->x.src0, coords->x.src1,
+ coords->x.dst0, coords->x.dst1,
+ coords->x.mirror);
brw_blorp_setup_coord_transform(¶ms->wm_inputs.coord_transform[1],
- src_y0, src_y1, dst_y0, dst_y1, mirror_y);
+ coords->y.src0, coords->y.src1,
+ coords->y.dst0, coords->y.dst1,
+ coords->y.mirror);
if (devinfo->gen > 6 &&
params->dst.surf.msaa_layout == ISL_MSAA_LAYOUT_INTERLEAVED) {
wm_prog_key->persample_msaa_dispatch = true;
}
+ params->num_samples = params->dst.surf.samples;
+
if (params->src.tile_x_sa || params->src.tile_y_sa) {
assert(wm_prog_key->need_src_offset);
surf_get_intratile_offset_px(¶ms->src,
/* For some texture types, we need to pass the layer through the sampler. */
params->wm_inputs.src_z = params->src.z_offset;
- brw_blorp_get_blit_kernel(batch->blorp, params, wm_prog_key);
+ if (!brw_blorp_get_blit_kernel(batch->blorp, params, wm_prog_key))
+ return 0;
+
+ unsigned result = 0;
+ unsigned max_surface_size = get_max_surface_size(devinfo, params);
+ if (params->src.surf.logical_level0_px.width > max_surface_size ||
+ params->dst.surf.logical_level0_px.width > max_surface_size)
+ result |= BLIT_WIDTH_SHRINK;
+ if (params->src.surf.logical_level0_px.height > max_surface_size ||
+ params->dst.surf.logical_level0_px.height > max_surface_size)
+ result |= BLIT_HEIGHT_SHRINK;
+
+ if (result == 0) {
+ batch->blorp->exec(batch, params);
+ }
+
+ return result;
+}
+
+/* Adjust split blit source coordinates for the current destination
+ * coordinates.
+ */
+static void
+adjust_split_source_coords(const struct blt_axis *orig,
+ struct blt_axis *split_coords,
+ double scale)
+{
+ /* When scale is greater than 0, then we are growing from the start, so
+ * src0 uses delta0, and src1 uses delta1. When scale is less than 0, the
+ * source range shrinks from the end. In that case src0 is adjusted by
+ * delta1, and src1 is adjusted by delta0.
+ */
+ double delta0 = scale * (split_coords->dst0 - orig->dst0);
+ double delta1 = scale * (split_coords->dst1 - orig->dst1);
+ split_coords->src0 = orig->src0 + (scale >= 0.0 ? delta0 : delta1);
+ split_coords->src1 = orig->src1 + (scale >= 0.0 ? delta1 : delta0);
+}
+
+static struct isl_extent2d
+get_px_size_sa(const struct isl_surf *surf)
+{
+ static const struct isl_extent2d one_to_one = { .w = 1, .h = 1 };
- batch->blorp->exec(batch, params);
+ if (surf->msaa_layout != ISL_MSAA_LAYOUT_INTERLEAVED)
+ return one_to_one;
+ else
+ return isl_get_interleaved_msaa_px_size_sa(surf->samples);
+}
+
+static void
+shrink_surface_params(const struct isl_device *dev,
+ struct brw_blorp_surface_info *info,
+ double *x0, double *x1, double *y0, double *y1)
+{
+ uint32_t byte_offset, x_offset_sa, y_offset_sa, size;
+ struct isl_extent2d px_size_sa;
+ int adjust;
+
+ surf_convert_to_single_slice(dev, info);
+
+ px_size_sa = get_px_size_sa(&info->surf);
+
+ /* Because this gets called after we lower compressed images, the tile
+ * offsets may be non-zero and we need to incorporate them in our
+ * calculations.
+ */
+ x_offset_sa = (uint32_t)*x0 * px_size_sa.w + info->tile_x_sa;
+ y_offset_sa = (uint32_t)*y0 * px_size_sa.h + info->tile_y_sa;
+ isl_tiling_get_intratile_offset_sa(dev, info->surf.tiling,
+ info->surf.format, info->surf.row_pitch,
+ x_offset_sa, y_offset_sa,
+ &byte_offset,
+ &info->tile_x_sa, &info->tile_y_sa);
+
+ info->addr.offset += byte_offset;
+
+ adjust = (int)info->tile_x_sa / px_size_sa.w - (int)*x0;
+ *x0 += adjust;
+ *x1 += adjust;
+ info->tile_x_sa = 0;
+
+ adjust = (int)info->tile_y_sa / px_size_sa.h - (int)*y0;
+ *y0 += adjust;
+ *y1 += adjust;
+ info->tile_y_sa = 0;
+
+ size = MIN2((uint32_t)ceil(*x1), info->surf.logical_level0_px.width);
+ info->surf.logical_level0_px.width = size;
+ info->surf.phys_level0_sa.width = size * px_size_sa.w;
+
+ size = MIN2((uint32_t)ceil(*y1), info->surf.logical_level0_px.height);
+ info->surf.logical_level0_px.height = size;
+ info->surf.phys_level0_sa.height = size * px_size_sa.h;
+}
+
+static void
+shrink_surfaces(const struct isl_device *dev,
+ struct blorp_params *params,
+ struct brw_blorp_blit_prog_key *wm_prog_key,
+ struct blt_coords *coords)
+{
+ /* Shrink source surface */
+ shrink_surface_params(dev, ¶ms->src, &coords->x.src0, &coords->x.src1,
+ &coords->y.src0, &coords->y.src1);
+ wm_prog_key->need_src_offset = false;
+
+ /* Shrink destination surface */
+ shrink_surface_params(dev, ¶ms->dst, &coords->x.dst0, &coords->x.dst1,
+ &coords->y.dst0, &coords->y.dst1);
+ wm_prog_key->need_dst_offset = false;
+}
+
+static void
+do_blorp_blit(struct blorp_batch *batch,
+ const struct blorp_params *orig_params,
+ struct brw_blorp_blit_prog_key *wm_prog_key,
+ const struct blt_coords *orig)
+{
+ struct blorp_params params;
+ struct blt_coords blit_coords;
+ struct blt_coords split_coords = *orig;
+ double w = orig->x.dst1 - orig->x.dst0;
+ double h = orig->y.dst1 - orig->y.dst0;
+ double x_scale = (orig->x.src1 - orig->x.src0) / w;
+ double y_scale = (orig->y.src1 - orig->y.src0) / h;
+ if (orig->x.mirror)
+ x_scale = -x_scale;
+ if (orig->y.mirror)
+ y_scale = -y_scale;
+
+ bool x_done, y_done;
+ bool shrink = split_blorp_blit_debug && can_shrink_surfaces(orig_params);
+ do {
+ params = *orig_params;
+ blit_coords = split_coords;
+ if (shrink)
+ shrink_surfaces(batch->blorp->isl_dev, ¶ms, wm_prog_key,
+ &blit_coords);
+ enum blit_shrink_status result =
+ try_blorp_blit(batch, ¶ms, wm_prog_key, &blit_coords);
+
+ if (result & BLIT_WIDTH_SHRINK) {
+ w /= 2.0;
+ assert(w >= 1.0);
+ split_coords.x.dst1 = MIN2(split_coords.x.dst0 + w, orig->x.dst1);
+ adjust_split_source_coords(&orig->x, &split_coords.x, x_scale);
+ }
+ if (result & BLIT_HEIGHT_SHRINK) {
+ h /= 2.0;
+ assert(h >= 1.0);
+ split_coords.y.dst1 = MIN2(split_coords.y.dst0 + h, orig->y.dst1);
+ adjust_split_source_coords(&orig->y, &split_coords.y, y_scale);
+ }
+
+ if (result != 0) {
+ assert(can_shrink_surfaces(orig_params));
+ shrink = true;
+ continue;
+ }
+
+ y_done = (orig->y.dst1 - split_coords.y.dst1 < 0.5);
+ x_done = y_done && (orig->x.dst1 - split_coords.x.dst1 < 0.5);
+ if (x_done) {
+ break;
+ } else if (y_done) {
+ split_coords.x.dst0 += w;
+ split_coords.x.dst1 = MIN2(split_coords.x.dst0 + w, orig->x.dst1);
+ split_coords.y.dst0 = orig->y.dst0;
+ split_coords.y.dst1 = MIN2(split_coords.y.dst0 + h, orig->y.dst1);
+ adjust_split_source_coords(&orig->x, &split_coords.x, x_scale);
+ } else {
+ split_coords.y.dst0 += h;
+ split_coords.y.dst1 = MIN2(split_coords.y.dst0 + h, orig->y.dst1);
+ adjust_split_source_coords(&orig->y, &split_coords.y, y_scale);
+ }
+ } while (true);
}
void
params.src.view.swizzle = src_swizzle;
params.dst.view.swizzle = dst_swizzle;
- struct brw_blorp_blit_prog_key wm_prog_key;
- memset(&wm_prog_key, 0, sizeof(wm_prog_key));
+ struct brw_blorp_blit_prog_key wm_prog_key = {
+ .shader_type = BLORP_SHADER_TYPE_BLIT
+ };
/* Scaled blitting or not. */
wm_prog_key.blit_scaled =
minify(params.src.surf.logical_level0_px.height, src_level) *
wm_prog_key.y_scale - 1.0f;
- do_blorp_blit(batch, ¶ms, &wm_prog_key,
- src_x0, src_y0, src_x1, src_y1,
- dst_x0, dst_y0, dst_x1, dst_y1,
- mirror_x, mirror_y);
+ struct blt_coords coords = {
+ .x = {
+ .src0 = src_x0,
+ .src1 = src_x1,
+ .dst0 = dst_x0,
+ .dst1 = dst_x1,
+ .mirror = mirror_x
+ },
+ .y = {
+ .src0 = src_y0,
+ .src1 = src_y1,
+ .dst0 = dst_y0,
+ .dst1 = dst_y1,
+ .mirror = mirror_y
+ }
+ };
+
+ do_blorp_blit(batch, ¶ms, &wm_prog_key, &coords);
}
static enum isl_format
-get_copy_format_for_bpb(unsigned bpb)
+get_copy_format_for_bpb(const struct isl_device *isl_dev, unsigned bpb)
{
- /* The choice of UNORM and UINT formats is very intentional here. Most of
- * the time, we want to use a UINT format to avoid any rounding error in
- * the blit. For stencil blits, R8_UINT is required by the hardware.
+ /* The choice of UNORM and UINT formats is very intentional here. Most
+ * of the time, we want to use a UINT format to avoid any rounding error
+ * in the blit. For stencil blits, R8_UINT is required by the hardware.
* (It's the only format allowed in conjunction with W-tiling.) Also we
* intentionally use the 4-channel formats whenever we can. This is so
- * that, when we do a RGB <-> RGBX copy, the two formats will line up even
- * though one of them is 3/4 the size of the other. The choice of UNORM
- * vs. UINT is also very intentional because Haswell doesn't handle 8 or
- * 16-bit RGB UINT formats at all so we have to use UNORM there.
+ * that, when we do a RGB <-> RGBX copy, the two formats will line up
+ * even though one of them is 3/4 the size of the other. The choice of
+ * UNORM vs. UINT is also very intentional because we don't have 8 or
+ * 16-bit RGB UINT formats until Sky Lake so we have to use UNORM there.
* Fortunately, the only time we should ever use two different formats in
* the table below is for RGB -> RGBA blits and so we will never have any
* UNORM/UINT mismatch.
*/
- switch (bpb) {
- case 8: return ISL_FORMAT_R8_UINT;
- case 16: return ISL_FORMAT_R8G8_UINT;
- case 24: return ISL_FORMAT_R8G8B8_UNORM;
- case 32: return ISL_FORMAT_R8G8B8A8_UNORM;
- case 48: return ISL_FORMAT_R16G16B16_UNORM;
- case 64: return ISL_FORMAT_R16G16B16A16_UNORM;
- case 96: return ISL_FORMAT_R32G32B32_UINT;
- case 128:return ISL_FORMAT_R32G32B32A32_UINT;
+ if (ISL_DEV_GEN(isl_dev) >= 9) {
+ switch (bpb) {
+ case 8: return ISL_FORMAT_R8_UINT;
+ case 16: return ISL_FORMAT_R8G8_UINT;
+ case 24: return ISL_FORMAT_R8G8B8_UINT;
+ case 32: return ISL_FORMAT_R8G8B8A8_UINT;
+ case 48: return ISL_FORMAT_R16G16B16_UINT;
+ case 64: return ISL_FORMAT_R16G16B16A16_UINT;
+ case 96: return ISL_FORMAT_R32G32B32_UINT;
+ case 128:return ISL_FORMAT_R32G32B32A32_UINT;
+ default:
+ unreachable("Unknown format bpb");
+ }
+ } else {
+ switch (bpb) {
+ case 8: return ISL_FORMAT_R8_UINT;
+ case 16: return ISL_FORMAT_R8G8_UINT;
+ case 24: return ISL_FORMAT_R8G8B8_UNORM;
+ case 32: return ISL_FORMAT_R8G8B8A8_UNORM;
+ case 48: return ISL_FORMAT_R16G16B16_UNORM;
+ case 64: return ISL_FORMAT_R16G16B16A16_UNORM;
+ case 96: return ISL_FORMAT_R32G32B32_UINT;
+ case 128:return ISL_FORMAT_R32G32B32A32_UINT;
+ default:
+ unreachable("Unknown format bpb");
+ }
+ }
+}
+
+/** Returns a UINT format that is CCS-compatible with the given format
+ *
+ * The PRM's say absolutely nothing about how render compression works. The
+ * only thing they provide is a list of formats on which it is and is not
+ * supported. Empirical testing indicates that the compression is only based
+ * on the bit-layout of the format and the channel encoding doesn't matter.
+ * So, while texture views don't work in general, you can create a view as
+ * long as the bit-layout of the formats are the same.
+ *
+ * Fortunately, for every render compression capable format, the UINT format
+ * with the same bit layout also supports render compression. This means that
+ * we only need to handle UINT formats for copy operations. In order to do
+ * copies between formats with different bit layouts, we attach both with a
+ * UINT format and use bit_cast_color() to generate code to do the bit-cast
+ * operation between the two bit layouts.
+ */
+static enum isl_format
+get_ccs_compatible_uint_format(const struct isl_format_layout *fmtl)
+{
+ switch (fmtl->format) {
+ case ISL_FORMAT_R32G32B32A32_FLOAT:
+ case ISL_FORMAT_R32G32B32A32_SINT:
+ case ISL_FORMAT_R32G32B32A32_UINT:
+ case ISL_FORMAT_R32G32B32A32_UNORM:
+ case ISL_FORMAT_R32G32B32A32_SNORM:
+ return ISL_FORMAT_R32G32B32A32_UINT;
+
+ case ISL_FORMAT_R16G16B16A16_UNORM:
+ case ISL_FORMAT_R16G16B16A16_SNORM:
+ case ISL_FORMAT_R16G16B16A16_SINT:
+ case ISL_FORMAT_R16G16B16A16_UINT:
+ case ISL_FORMAT_R16G16B16A16_FLOAT:
+ case ISL_FORMAT_R16G16B16X16_UNORM:
+ case ISL_FORMAT_R16G16B16X16_FLOAT:
+ return ISL_FORMAT_R16G16B16A16_UINT;
+
+ case ISL_FORMAT_R32G32_FLOAT:
+ case ISL_FORMAT_R32G32_SINT:
+ case ISL_FORMAT_R32G32_UINT:
+ case ISL_FORMAT_R32G32_UNORM:
+ case ISL_FORMAT_R32G32_SNORM:
+ return ISL_FORMAT_R32G32_UINT;
+
+ case ISL_FORMAT_B8G8R8A8_UNORM:
+ case ISL_FORMAT_B8G8R8A8_UNORM_SRGB:
+ case ISL_FORMAT_R8G8B8A8_UNORM:
+ case ISL_FORMAT_R8G8B8A8_UNORM_SRGB:
+ case ISL_FORMAT_R8G8B8A8_SNORM:
+ case ISL_FORMAT_R8G8B8A8_SINT:
+ case ISL_FORMAT_R8G8B8A8_UINT:
+ case ISL_FORMAT_B8G8R8X8_UNORM:
+ case ISL_FORMAT_B8G8R8X8_UNORM_SRGB:
+ case ISL_FORMAT_R8G8B8X8_UNORM:
+ case ISL_FORMAT_R8G8B8X8_UNORM_SRGB:
+ return ISL_FORMAT_R8G8B8A8_UINT;
+
+ case ISL_FORMAT_R16G16_UNORM:
+ case ISL_FORMAT_R16G16_SNORM:
+ case ISL_FORMAT_R16G16_SINT:
+ case ISL_FORMAT_R16G16_UINT:
+ case ISL_FORMAT_R16G16_FLOAT:
+ return ISL_FORMAT_R16G16_UINT;
+
+ case ISL_FORMAT_R32_SINT:
+ case ISL_FORMAT_R32_UINT:
+ case ISL_FORMAT_R32_FLOAT:
+ case ISL_FORMAT_R32_UNORM:
+ case ISL_FORMAT_R32_SNORM:
+ return ISL_FORMAT_R32_UINT;
+
+ default:
+ unreachable("Not a compressible format");
+ }
+}
+
+/* Takes an isl_color_value and returns a color value that is the original
+ * color value only bit-casted to a UINT format. This value, together with
+ * the format from get_ccs_compatible_uint_format, will yield the same bit
+ * value as the original color and format.
+ */
+static union isl_color_value
+bitcast_color_value_to_uint(union isl_color_value color,
+ const struct isl_format_layout *fmtl)
+{
+ /* All CCS formats have the same number of bits in each channel */
+ const struct isl_channel_layout *chan = &fmtl->channels.r;
+
+ union isl_color_value bits;
+ switch (chan->type) {
+ case ISL_UINT:
+ case ISL_SINT:
+ /* Hardware will ignore the high bits so there's no need to cast */
+ bits = color;
+ break;
+
+ case ISL_UNORM:
+ for (unsigned i = 0; i < 4; i++)
+ bits.u32[i] = _mesa_float_to_unorm(color.f32[i], chan->bits);
+ break;
+
+ case ISL_SNORM:
+ for (unsigned i = 0; i < 4; i++)
+ bits.i32[i] = _mesa_float_to_snorm(color.f32[i], chan->bits);
+ break;
+
+ case ISL_SFLOAT:
+ switch (chan->bits) {
+ case 16:
+ for (unsigned i = 0; i < 4; i++)
+ bits.u32[i] = _mesa_float_to_half(color.f32[i]);
+ break;
+
+ case 32:
+ bits = color;
+ break;
+
+ default:
+ unreachable("Invalid float format size");
+ }
+ break;
+
+ default:
+ unreachable("Invalid channel type");
+ }
+
+ switch (fmtl->format) {
+ case ISL_FORMAT_B8G8R8A8_UNORM:
+ case ISL_FORMAT_B8G8R8A8_UNORM_SRGB:
+ case ISL_FORMAT_B8G8R8X8_UNORM:
+ case ISL_FORMAT_B8G8R8X8_UNORM_SRGB: {
+ /* If it's a BGRA format, we need to swap blue and red */
+ uint32_t tmp = bits.u32[0];
+ bits.u32[0] = bits.u32[2];
+ bits.u32[2] = tmp;
+ break;
+ }
+
default:
- unreachable("Unknown format bpb");
+ break; /* Nothing to do */
}
+
+ return bits;
}
static void
surf_convert_to_single_slice(isl_dev, info);
if (width || height) {
+#ifndef NDEBUG
+ uint32_t right_edge_px = info->tile_x_sa + *x + *width;
+ uint32_t bottom_edge_px = info->tile_y_sa + *y + *height;
assert(*width % fmtl->bw == 0 ||
- *x + *width == info->surf.logical_level0_px.width);
+ right_edge_px == info->surf.logical_level0_px.width);
assert(*height % fmtl->bh == 0 ||
- *y + *height == info->surf.logical_level0_px.height);
+ bottom_edge_px == info->surf.logical_level0_px.height);
+#endif
*width = DIV_ROUND_UP(*width, fmtl->bw);
*height = DIV_ROUND_UP(*height, fmtl->bh);
}
info->tile_y_sa /= fmtl->bh;
/* It's now an uncompressed surface so we need an uncompressed format */
- info->surf.format = get_copy_format_for_bpb(fmtl->bpb);
+ info->surf.format = get_copy_format_for_bpb(isl_dev, fmtl->bpb);
}
void
uint32_t dst_x, uint32_t dst_y,
uint32_t src_width, uint32_t src_height)
{
+ const struct isl_device *isl_dev = batch->blorp->isl_dev;
struct blorp_params params;
- blorp_params_init(¶ms);
+ if (src_width == 0 || src_height == 0)
+ return;
+
+ blorp_params_init(¶ms);
brw_blorp_surface_info_init(batch->blorp, ¶ms.src, src_surf, src_level,
src_layer, ISL_FORMAT_UNSUPPORTED, false);
brw_blorp_surface_info_init(batch->blorp, ¶ms.dst, dst_surf, dst_level,
dst_layer, ISL_FORMAT_UNSUPPORTED, true);
- struct brw_blorp_blit_prog_key wm_prog_key;
- memset(&wm_prog_key, 0, sizeof(wm_prog_key));
+ struct brw_blorp_blit_prog_key wm_prog_key = {
+ .shader_type = BLORP_SHADER_TYPE_BLIT
+ };
const struct isl_format_layout *src_fmtl =
isl_format_get_layout(params.src.surf.format);
const struct isl_format_layout *dst_fmtl =
isl_format_get_layout(params.dst.surf.format);
- params.src.view.format = get_copy_format_for_bpb(src_fmtl->bpb);
+ assert(params.src.aux_usage == ISL_AUX_USAGE_NONE ||
+ params.src.aux_usage == ISL_AUX_USAGE_MCS ||
+ params.src.aux_usage == ISL_AUX_USAGE_CCS_E);
+ assert(params.dst.aux_usage == ISL_AUX_USAGE_NONE ||
+ params.dst.aux_usage == ISL_AUX_USAGE_MCS ||
+ params.dst.aux_usage == ISL_AUX_USAGE_CCS_E);
+
+ if (params.dst.aux_usage == ISL_AUX_USAGE_CCS_E) {
+ params.dst.view.format = get_ccs_compatible_uint_format(dst_fmtl);
+ if (params.src.aux_usage == ISL_AUX_USAGE_CCS_E) {
+ params.src.view.format = get_ccs_compatible_uint_format(src_fmtl);
+ } else if (src_fmtl->bpb == dst_fmtl->bpb) {
+ params.src.view.format = params.dst.view.format;
+ } else {
+ params.src.view.format =
+ get_copy_format_for_bpb(isl_dev, src_fmtl->bpb);
+ }
+ } else if (params.src.aux_usage == ISL_AUX_USAGE_CCS_E) {
+ params.src.view.format = get_ccs_compatible_uint_format(src_fmtl);
+ if (src_fmtl->bpb == dst_fmtl->bpb) {
+ params.dst.view.format = params.src.view.format;
+ } else {
+ params.dst.view.format =
+ get_copy_format_for_bpb(isl_dev, dst_fmtl->bpb);
+ }
+ } else {
+ params.dst.view.format = get_copy_format_for_bpb(isl_dev, dst_fmtl->bpb);
+ params.src.view.format = get_copy_format_for_bpb(isl_dev, src_fmtl->bpb);
+ }
+
+ if (params.src.aux_usage == ISL_AUX_USAGE_CCS_E) {
+ assert(isl_formats_are_ccs_e_compatible(batch->blorp->isl_dev->info,
+ src_surf->surf->format,
+ params.src.view.format));
+ params.src.clear_color =
+ bitcast_color_value_to_uint(params.src.clear_color, src_fmtl);
+ }
+
+ if (params.dst.aux_usage == ISL_AUX_USAGE_CCS_E) {
+ assert(isl_formats_are_ccs_e_compatible(batch->blorp->isl_dev->info,
+ dst_surf->surf->format,
+ params.dst.view.format));
+ params.dst.clear_color =
+ bitcast_color_value_to_uint(params.dst.clear_color, dst_fmtl);
+ }
+
+ wm_prog_key.src_bpc =
+ isl_format_get_layout(params.src.view.format)->channels.r.bits;
+ wm_prog_key.dst_bpc =
+ isl_format_get_layout(params.dst.view.format)->channels.r.bits;
+
if (src_fmtl->bw > 1 || src_fmtl->bh > 1) {
surf_convert_to_uncompressed(batch->blorp->isl_dev, ¶ms.src,
&src_x, &src_y, &src_width, &src_height);
wm_prog_key.need_src_offset = true;
}
- params.dst.view.format = get_copy_format_for_bpb(dst_fmtl->bpb);
if (dst_fmtl->bw > 1 || dst_fmtl->bh > 1) {
surf_convert_to_uncompressed(batch->blorp->isl_dev, ¶ms.dst,
&dst_x, &dst_y, NULL, NULL);
uint32_t dst_width = src_width;
uint32_t dst_height = src_height;
- do_blorp_blit(batch, ¶ms, &wm_prog_key,
- src_x, src_y, src_x + src_width, src_y + src_height,
- dst_x, dst_y, dst_x + dst_width, dst_y + dst_height,
- false, false);
+ struct blt_coords coords = {
+ .x = {
+ .src0 = src_x,
+ .src1 = src_x + src_width,
+ .dst0 = dst_x,
+ .dst1 = dst_x + dst_width,
+ .mirror = false
+ },
+ .y = {
+ .src0 = src_y,
+ .src1 = src_y + src_height,
+ .dst0 = dst_y,
+ .dst1 = dst_y + dst_height,
+ .mirror = false
+ }
+ };
+
+ do_blorp_blit(batch, ¶ms, &wm_prog_key, &coords);
}