summaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers/radeonsi
diff options
context:
space:
mode:
authorMarek Olšák <[email protected]>2016-01-11 23:51:39 +0100
committerMarek Olšák <[email protected]>2016-01-22 15:02:40 +0100
commit8667a1aea2ac3fee9b1b663132dc1bea8ec21cda (patch)
treeec989c6f1677d180846bf894ee4849bd83af3040 /src/gallium/drivers/radeonsi
parent0446ea9d084f5e72f7ea23f806a9723c94dd3685 (diff)
radeonsi: use SPI_SHADER_COL_FORMAT fields instead of export_16bpc
This does change the behavior slightly: If a shader writes COLOR[i] and that color buffer isn't bound, the shader will export MRT_NULL instead and discard the IR tree that calculates the output. The only exception is alpha-to-coverage, which requires an alpha export. v2: - update a comment about 16BPC - account for MRTZ when when fixing alpha-test/kill Reviewed-by: Nicolai Hähnle <[email protected]>
Diffstat (limited to 'src/gallium/drivers/radeonsi')
-rw-r--r--src/gallium/drivers/radeonsi/si_pipe.h2
-rw-r--r--src/gallium/drivers/radeonsi/si_shader.c35
-rw-r--r--src/gallium/drivers/radeonsi/si_shader.h2
-rw-r--r--src/gallium/drivers/radeonsi/si_state.c41
-rw-r--r--src/gallium/drivers/radeonsi/si_state.h1
-rw-r--r--src/gallium/drivers/radeonsi/si_state_shaders.c47
6 files changed, 90 insertions, 38 deletions
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index f83cb024f0e..e2009de9870 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -125,7 +125,7 @@ struct si_framebuffer {
unsigned log_samples;
unsigned cb0_is_integer;
unsigned compressed_cb_mask;
- unsigned export_16bpc;
+ unsigned spi_shader_col_format;
unsigned dirty_cbufs;
bool dirty_zsbuf;
};
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 2de7def8dd2..266ef6db633 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1265,7 +1265,7 @@ static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
struct lp_build_context *uint =
&si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
struct lp_build_context *base = &bld_base->base;
- unsigned compressed = 0;
+ unsigned spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
unsigned chan;
/* XXX: This controls which components of the output
@@ -1286,17 +1286,29 @@ static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
args[3] = lp_build_const_int32(base->gallivm, target);
if (si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT) {
+ unsigned col_formats =
+ si_shader_ctx->shader->key.ps.spi_shader_col_format;
int cbuf = target - V_008DFC_SQ_EXP_MRT;
- if (cbuf >= 0 && cbuf < 8)
- compressed = (si_shader_ctx->shader->key.ps.export_16bpc >> cbuf) & 0x1;
+ assert(cbuf >= 0 && cbuf < 8);
+ spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
}
- /* Set COMPR flag */
- args[4] = compressed ? uint->one : uint->zero;
+ args[4] = uint->zero; /* COMPR flag */
+ args[5] = base->undef;
+ args[6] = base->undef;
+ args[7] = base->undef;
+ args[8] = base->undef;
+
+ switch (spi_shader_col_format) {
+ case V_028714_SPI_SHADER_ZERO:
+ args[0] = uint->zero; /* writemask */
+ args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_NULL);
+ break;
+
+ case V_028714_SPI_SHADER_FP16_ABGR:
+ args[4] = uint->one; /* COMPR flag */
- if (compressed) {
- /* Pixel shader needs to pack output values before export */
for (chan = 0; chan < 2; chan++) {
LLVMValueRef pack_args[2] = {
values[2 * chan],
@@ -1314,10 +1326,13 @@ static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
packed,
LLVMFloatTypeInContext(base->gallivm->context),
"");
- args[chan + 7] = base->undef;
}
- } else
+ break;
+
+ case V_028714_SPI_SHADER_32_ABGR:
memcpy(&args[5], values, sizeof(values[0]) * 4);
+ break;
+ }
}
static void si_alpha_test(struct lp_build_tgsi_context *bld_base,
@@ -4031,7 +4046,7 @@ void si_dump_shader_key(unsigned shader, union si_shader_key *key, FILE *f)
break;
case PIPE_SHADER_FRAGMENT:
- fprintf(f, " export_16bpc = 0x%X\n", key->ps.export_16bpc);
+ fprintf(f, " spi_shader_col_format = 0x%x\n", key->ps.spi_shader_col_format);
fprintf(f, " last_cbuf = %u\n", key->ps.last_cbuf);
fprintf(f, " color_two_side = %u\n", key->ps.color_two_side);
fprintf(f, " alpha_func = %u\n", key->ps.alpha_func);
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index 1635358d505..a9b76c6d097 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -232,7 +232,7 @@ struct si_shader_selector {
union si_shader_key {
struct {
- unsigned export_16bpc:8;
+ unsigned spi_shader_col_format;
unsigned last_cbuf:3;
unsigned color_two_side:1;
unsigned alpha_func:3;
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index ea441acb1b0..af6759ef19c 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -403,6 +403,7 @@ static void *si_create_blend_state_mode(struct pipe_context *ctx,
if (!blend)
return NULL;
+ blend->alpha_to_coverage = state->alpha_to_coverage;
blend->alpha_to_one = state->alpha_to_one;
blend->dual_src_blend = util_blend_state_is_dual(state, 0);
@@ -1883,6 +1884,21 @@ unsigned si_tile_mode_index(struct r600_texture *rtex, unsigned level, bool sten
* framebuffer handling
*/
+static void si_choose_spi_color_formats(struct r600_surface *surf,
+ unsigned format, unsigned swap,
+ unsigned ntype)
+{
+ unsigned max_comp_size = si_colorformat_max_comp_size(format);
+
+ surf->spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
+
+ if (ntype == V_028C70_NUMBER_SRGB ||
+ ((ntype == V_028C70_NUMBER_UNORM || ntype == V_028C70_NUMBER_SNORM) &&
+ max_comp_size <= 10) ||
+ (ntype == V_028C70_NUMBER_FLOAT && max_comp_size <= 16))
+ surf->spi_shader_col_format = V_028714_SPI_SHADER_FP16_ABGR;
+}
+
static void si_initialize_color_surface(struct si_context *sctx,
struct r600_surface *surf)
{
@@ -1896,7 +1912,6 @@ static void si_initialize_color_surface(struct si_context *sctx,
const struct util_format_description *desc;
int i;
unsigned blend_clamp = 0, blend_bypass = 0;
- unsigned max_comp_size;
/* Layered rendering doesn't work with LINEAR_GENERAL.
* (LINEAR_ALIGNED and others work) */
@@ -2053,13 +2068,7 @@ static void si_initialize_color_surface(struct si_context *sctx,
}
/* Determine pixel shader export format */
- max_comp_size = si_colorformat_max_comp_size(format);
- if (ntype == V_028C70_NUMBER_SRGB ||
- ((ntype == V_028C70_NUMBER_UNORM || ntype == V_028C70_NUMBER_SNORM) &&
- max_comp_size <= 10) ||
- (ntype == V_028C70_NUMBER_FLOAT && max_comp_size <= 16)) {
- surf->export_16bpc = true;
- }
+ si_choose_spi_color_formats(surf, format, swap, ntype);
if (sctx->b.family == CHIP_STONEY &&
!(sctx->screen->b.debug_flags & DBG_NO_RB_PLUS)) {
@@ -2286,7 +2295,7 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
util_copy_framebuffer_state(&sctx->framebuffer.state, state);
- sctx->framebuffer.export_16bpc = 0;
+ sctx->framebuffer.spi_shader_col_format = 0;
sctx->framebuffer.compressed_cb_mask = 0;
sctx->framebuffer.nr_samples = util_framebuffer_get_num_samples(state);
sctx->framebuffer.log_samples = util_logbase2(sctx->framebuffer.nr_samples);
@@ -2307,22 +2316,20 @@ static void si_set_framebuffer_state(struct pipe_context *ctx,
si_initialize_color_surface(sctx, surf);
}
- if (surf->export_16bpc) {
- sctx->framebuffer.export_16bpc |= 1 << i;
- }
+ sctx->framebuffer.spi_shader_col_format |=
+ surf->spi_shader_col_format << (i * 4);
if (rtex->fmask.size && rtex->cmask.size) {
sctx->framebuffer.compressed_cb_mask |= 1 << i;
}
r600_context_add_resource_size(ctx, surf->base.texture);
}
- /* Set the 16BPC export for possible dual-src blending. */
- if (i == 1 && surf && surf->export_16bpc) {
- sctx->framebuffer.export_16bpc |= 1 << 1;
+ /* Set the second SPI format for possible dual-src blending. */
+ if (i == 1 && surf) {
+ sctx->framebuffer.spi_shader_col_format |=
+ surf->spi_shader_col_format << (i * 4);
}
- assert(!(sctx->framebuffer.export_16bpc & ~0xff));
-
if (state->zsbuf) {
surf = (struct r600_surface*)state->zsbuf;
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index f5ca661f8d7..46ba3c4301f 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -39,6 +39,7 @@ struct si_shader;
struct si_state_blend {
struct si_pm4_state pm4;
uint32_t cb_target_mask;
+ bool alpha_to_coverage;
bool alpha_to_one;
bool dual_src_blend;
};
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 8ff70b44d45..91ba9b76210 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -382,13 +382,28 @@ static void si_shader_vs(struct si_shader *shader)
si_set_tesseval_regs(shader, pm4);
}
+static unsigned si_get_spi_shader_col_format(struct si_shader *shader)
+{
+ unsigned value = shader->key.ps.spi_shader_col_format;
+ unsigned i, num_targets = (util_last_bit(value) + 3) / 4;
+
+ /* If the i-th target format is set, all previous target formats must
+ * be non-zero to avoid hangs.
+ */
+ for (i = 0; i < num_targets; i++)
+ if (!(value & (0xf << (i * 4))))
+ value |= V_028714_SPI_SHADER_32_R << (i * 4);
+
+ return value;
+}
+
static void si_shader_ps(struct si_shader *shader)
{
struct tgsi_shader_info *info = &shader->selector->info;
struct si_pm4_state *pm4;
- unsigned i, spi_ps_in_control;
- unsigned spi_shader_col_format = 0, cb_shader_mask = 0;
- unsigned colors_written, export_16bpc;
+ unsigned i, spi_ps_in_control, spi_shader_col_format;
+ unsigned cb_shader_mask = 0;
+ unsigned colors_written;
unsigned num_sgprs, num_user_sgprs;
unsigned spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
uint64_t va;
@@ -425,7 +440,6 @@ static void si_shader_ps(struct si_shader *shader)
/* Find out what SPI_SHADER_COL_FORMAT and CB_SHADER_MASK should be. */
colors_written = info->colors_written;
- export_16bpc = shader->key.ps.export_16bpc;
if (info->colors_written == 0x1 &&
info->properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS]) {
@@ -434,13 +448,20 @@ static void si_shader_ps(struct si_shader *shader)
while (colors_written) {
i = u_bit_scan(&colors_written);
- if (export_16bpc & (1 << i))
- spi_shader_col_format |= V_028714_SPI_SHADER_FP16_ABGR << (4 * i);
- else
- spi_shader_col_format |= V_028714_SPI_SHADER_32_ABGR << (4 * i);
cb_shader_mask |= 0xf << (4 * i);
}
+ spi_shader_col_format = si_get_spi_shader_col_format(shader);
+
+ /* This must be non-zero for alpha-test/kill to work.
+ * The hardware ignores the EXEC mask if no export memory is allocated.
+ */
+ if (!spi_shader_col_format &&
+ !info->writes_z && !info->writes_stencil && !info->writes_samplemask &&
+ (shader->selector->info.uses_kill ||
+ shader->key.ps.alpha_func != PIPE_FUNC_ALWAYS))
+ spi_shader_col_format = V_028714_SPI_SHADER_32_R;
+
/* Set interpolation controls. */
has_centroid = G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena) ||
G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena);
@@ -571,12 +592,20 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
break;
case PIPE_SHADER_FRAGMENT: {
struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
+ struct si_state_blend *blend = sctx->queued.named.blend;
if (sel->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
sel->info.colors_written == 0x1)
key->ps.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
- key->ps.export_16bpc = sctx->framebuffer.export_16bpc;
+ key->ps.spi_shader_col_format = sctx->framebuffer.spi_shader_col_format;
+
+ /* If alpha-to-coverage is enabled, we have to export alpha
+ * even if there is no color buffer.
+ */
+ if (!(key->ps.spi_shader_col_format & 0xf) &&
+ blend && blend->alpha_to_coverage)
+ key->ps.spi_shader_col_format |= V_028710_SPI_SHADER_FP16_ABGR;
if (rs) {
bool is_poly = (sctx->current_rast_prim >= PIPE_PRIM_TRIANGLES &&