summaryrefslogtreecommitdiffstats
path: root/src
diff options
context:
space:
mode:
authorDave Airlie <[email protected]>2014-11-24 12:53:40 +1000
committerDave Airlie <[email protected]>2014-11-27 10:31:38 +1000
commit07ae69753c6818bcce5d4edaf2cca39c20e37f4c (patch)
tree94119bf2ec201d83c630943cc0921dabe5de2150 /src
parentbc5f5424e3a8cf0cd0a07bcf3d85670735a49f41 (diff)
r600g: merge the TXQ and BUFFER constant buffers (v1.1)
We are using 1 more buffer than we have, although in the future the driver should just end up using one buffer in total probably, this is a good first step, it merges the txq cube array and buffer info constants on r600 and evergreen. This should in theory fix geom shader tests on r600. v1.1: fix comments from Glenn. Reviewed-by: Glenn Kennard <[email protected]> Cc: "10.4 10.3" <[email protected] Signed-off-by: Dave Airlie <[email protected]>
Diffstat (limited to 'src')
-rw-r--r--src/gallium/drivers/r600/r600_pipe.h15
-rw-r--r--src/gallium/drivers/r600/r600_shader.c18
-rw-r--r--src/gallium/drivers/r600/r600_state_common.c96
3 files changed, 61 insertions, 68 deletions
diff --git a/src/gallium/drivers/r600/r600_pipe.h b/src/gallium/drivers/r600/r600_pipe.h
index 40b0328ea20..da959dd537b 100644
--- a/src/gallium/drivers/r600/r600_pipe.h
+++ b/src/gallium/drivers/r600/r600_pipe.h
@@ -44,20 +44,19 @@
#define R600_TRACE_CS_DWORDS 7
#define R600_MAX_USER_CONST_BUFFERS 13
-#define R600_MAX_DRIVER_CONST_BUFFERS 4
+#define R600_MAX_DRIVER_CONST_BUFFERS 3
#define R600_MAX_CONST_BUFFERS (R600_MAX_USER_CONST_BUFFERS + R600_MAX_DRIVER_CONST_BUFFERS)
/* start driver buffers after user buffers */
#define R600_UCP_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS)
-#define R600_TXQ_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS + 1)
-#define R600_BUFFER_INFO_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS + 2)
-#define R600_GS_RING_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS + 3)
-/* Currently R600_MAX_CONST_BUFFERS is too large, the hardware only has 16 buffers, but the driver is
- * trying to use 17. Avoid accidentally aliasing with user UBOs for SAMPLE_POSITIONS by using an id<16.
+#define R600_BUFFER_INFO_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS + 1)
+#define R600_GS_RING_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS + 2)
+/* Currently R600_MAX_CONST_BUFFERS just fits on the hw, which has a limit
+ * of 16 const buffers.
* UCP/SAMPLE_POSITIONS are never accessed by same shader stage so they can use the same id.
*
- * Fixing this properly would require the driver to combine its buffers into a single hardware buffer,
- * which would also allow supporting the d3d 11 mandated minimum of 15 user const buffers.
+ * In order to support d3d 11 mandated minimum of 15 user const buffers
+ * we'd have to squash all use cases into one driver buffer.
*/
#define R600_SAMPLE_POSITIONS_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS)
diff --git a/src/gallium/drivers/r600/r600_shader.c b/src/gallium/drivers/r600/r600_shader.c
index 7a2609a2c1d..7bb98728a99 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -5035,8 +5035,9 @@ static int r600_do_buffer_txq(struct r600_shader_ctx *ctx)
alu.op = ALU_OP1_MOV;
if (ctx->bc->chip_class >= EVERGREEN) {
- alu.src[0].sel = 512 + (id / 4);
- alu.src[0].chan = id % 4;
+ /* channel 0 or 2 of each word */
+ alu.src[0].sel = 512 + (id / 2);
+ alu.src[0].chan = (id % 2) * 2;
} else {
/* r600 we have them at channel 2 of the second dword */
alu.src[0].sel = 512 + (id * 2) + 1;
@@ -5697,9 +5698,16 @@ static int tgsi_tex(struct r600_shader_ctx *ctx)
memset(&alu, 0, sizeof(struct r600_bytecode_alu));
alu.op = ALU_OP1_MOV;
- alu.src[0].sel = 512 + (id / 4);
- alu.src[0].kc_bank = R600_TXQ_CONST_BUFFER;
- alu.src[0].chan = id % 4;
+ if (ctx->bc->chip_class >= EVERGREEN) {
+ /* channel 1 or 3 of each word */
+ alu.src[0].sel = 512 + (id / 2);
+ alu.src[0].chan = ((id % 2) * 2) + 1;
+ } else {
+ /* r600 we have them at channel 2 of the second dword */
+ alu.src[0].sel = 512 + (id * 2) + 1;
+ alu.src[0].chan = 2;
+ }
+ alu.src[0].kc_bank = R600_BUFFER_INFO_CONST_BUFFER;
tgsi_dst(ctx, &inst->Dst[0], 2, &alu.dst);
alu.last = 1;
r = r600_bytecode_add_alu(ctx->bc, &alu);
diff --git a/src/gallium/drivers/r600/r600_state_common.c b/src/gallium/drivers/r600/r600_state_common.c
index c3f21cb93ae..d35c9b363ef 100644
--- a/src/gallium/drivers/r600/r600_state_common.c
+++ b/src/gallium/drivers/r600/r600_state_common.c
@@ -984,6 +984,7 @@ static void r600_set_sample_mask(struct pipe_context *pipe, unsigned sample_mask
* then in the shader, we AND the 4 components with 0xffffffff or 0,
* then OR the alpha with the value given here.
* We use a 6th constant to store the txq buffer size in
+ * we use 7th slot for number of cube layers in a cube map array.
*/
static void r600_setup_buffer_constants(struct r600_context *rctx, int shader_type)
{
@@ -1022,6 +1023,7 @@ static void r600_setup_buffer_constants(struct r600_context *rctx, int shader_ty
samplers->buffer_constants[offset + 4] = 0;
samplers->buffer_constants[offset + 5] = samplers->views.views[i]->base.texture->width0 / util_format_get_blocksize(samplers->views.views[i]->base.format);
+ samplers->buffer_constants[offset + 6] = samplers->views.views[i]->base.texture->array_size / 6;
}
}
@@ -1033,7 +1035,10 @@ static void r600_setup_buffer_constants(struct r600_context *rctx, int shader_ty
pipe_resource_reference(&cb.buffer, NULL);
}
-/* On evergreen we only need to store the buffer size for TXQ */
+/* On evergreen we store two values
+ * 1. buffer size for TXQ
+ * 2. number of cube layers in a cube map array.
+ */
static void eg_setup_buffer_constants(struct r600_context *rctx, int shader_type)
{
struct r600_textures_info *samplers = &rctx->samplers[shader_type];
@@ -1048,12 +1053,16 @@ static void eg_setup_buffer_constants(struct r600_context *rctx, int shader_type
samplers->views.dirty_buffer_constants = FALSE;
bits = util_last_bit(samplers->views.enabled_mask);
- array_size = bits * sizeof(uint32_t) * 4;
+ array_size = bits * 2 * sizeof(uint32_t) * 4;
samplers->buffer_constants = realloc(samplers->buffer_constants, array_size);
memset(samplers->buffer_constants, 0, array_size);
- for (i = 0; i < bits; i++)
- if (samplers->views.enabled_mask & (1 << i))
- samplers->buffer_constants[i] = samplers->views.views[i]->base.texture->width0 / util_format_get_blocksize(samplers->views.views[i]->base.format);
+ for (i = 0; i < bits; i++) {
+ if (samplers->views.enabled_mask & (1 << i)) {
+ uint32_t offset = i * 2;
+ samplers->buffer_constants[offset] = samplers->views.views[i]->base.texture->width0 / util_format_get_blocksize(samplers->views.views[i]->base.format);
+ samplers->buffer_constants[offset + 1] = samplers->views.views[i]->base.texture->array_size / 6;
+ }
+ }
cb.buffer = NULL;
cb.user_buffer = samplers->buffer_constants;
@@ -1063,35 +1072,6 @@ static void eg_setup_buffer_constants(struct r600_context *rctx, int shader_type
pipe_resource_reference(&cb.buffer, NULL);
}
-static void r600_setup_txq_cube_array_constants(struct r600_context *rctx, int shader_type)
-{
- struct r600_textures_info *samplers = &rctx->samplers[shader_type];
- int bits;
- uint32_t array_size;
- struct pipe_constant_buffer cb;
- int i;
-
- if (!samplers->views.dirty_txq_constants)
- return;
-
- samplers->views.dirty_txq_constants = FALSE;
-
- bits = util_last_bit(samplers->views.enabled_mask);
- array_size = bits * sizeof(uint32_t) * 4;
- samplers->txq_constants = realloc(samplers->txq_constants, array_size);
- memset(samplers->txq_constants, 0, array_size);
- for (i = 0; i < bits; i++)
- if (samplers->views.enabled_mask & (1 << i))
- samplers->txq_constants[i] = samplers->views.views[i]->base.texture->array_size / 6;
-
- cb.buffer = NULL;
- cb.user_buffer = samplers->txq_constants;
- cb.buffer_offset = 0;
- cb.buffer_size = array_size;
- rctx->b.b.set_constant_buffer(&rctx->b.b, shader_type, R600_TXQ_CONST_BUFFER, &cb);
- pipe_resource_reference(&cb.buffer, NULL);
-}
-
/* set sample xy locations as array of fragment shader constants */
void r600_set_sample_locations_constant_buffer(struct r600_context *rctx)
{
@@ -1175,7 +1155,7 @@ static bool r600_update_derived_state(struct r600_context *rctx)
struct pipe_context * ctx = (struct pipe_context*)rctx;
bool ps_dirty = false, vs_dirty = false, gs_dirty = false;
bool blend_disable;
-
+ bool need_buf_const;
if (!rctx->blitter->running) {
unsigned i;
@@ -1296,29 +1276,35 @@ static bool r600_update_derived_state(struct r600_context *rctx)
/* on R600 we stuff masks + txq info into one constant buffer */
/* on evergreen we only need a txq info one */
- if (rctx->b.chip_class < EVERGREEN) {
- if (rctx->ps_shader && rctx->ps_shader->current->shader.uses_tex_buffers)
- r600_setup_buffer_constants(rctx, PIPE_SHADER_FRAGMENT);
- if (rctx->vs_shader && rctx->vs_shader->current->shader.uses_tex_buffers)
- r600_setup_buffer_constants(rctx, PIPE_SHADER_VERTEX);
- if (rctx->gs_shader && rctx->gs_shader->current->shader.uses_tex_buffers)
- r600_setup_buffer_constants(rctx, PIPE_SHADER_GEOMETRY);
- } else {
- if (rctx->ps_shader && rctx->ps_shader->current->shader.uses_tex_buffers)
- eg_setup_buffer_constants(rctx, PIPE_SHADER_FRAGMENT);
- if (rctx->vs_shader && rctx->vs_shader->current->shader.uses_tex_buffers)
- eg_setup_buffer_constants(rctx, PIPE_SHADER_VERTEX);
- if (rctx->gs_shader && rctx->gs_shader->current->shader.uses_tex_buffers)
- eg_setup_buffer_constants(rctx, PIPE_SHADER_GEOMETRY);
+ if (rctx->ps_shader) {
+ need_buf_const = rctx->ps_shader->current->shader.uses_tex_buffers || rctx->ps_shader->current->shader.has_txq_cube_array_z_comp;
+ if (need_buf_const) {
+ if (rctx->b.chip_class < EVERGREEN)
+ r600_setup_buffer_constants(rctx, PIPE_SHADER_FRAGMENT);
+ else
+ eg_setup_buffer_constants(rctx, PIPE_SHADER_FRAGMENT);
+ }
}
+ if (rctx->vs_shader) {
+ need_buf_const = rctx->vs_shader->current->shader.uses_tex_buffers || rctx->vs_shader->current->shader.has_txq_cube_array_z_comp;
+ if (need_buf_const) {
+ if (rctx->b.chip_class < EVERGREEN)
+ r600_setup_buffer_constants(rctx, PIPE_SHADER_VERTEX);
+ else
+ eg_setup_buffer_constants(rctx, PIPE_SHADER_VERTEX);
+ }
+ }
- if (rctx->ps_shader && rctx->ps_shader->current->shader.has_txq_cube_array_z_comp)
- r600_setup_txq_cube_array_constants(rctx, PIPE_SHADER_FRAGMENT);
- if (rctx->vs_shader && rctx->vs_shader->current->shader.has_txq_cube_array_z_comp)
- r600_setup_txq_cube_array_constants(rctx, PIPE_SHADER_VERTEX);
- if (rctx->gs_shader && rctx->gs_shader->current->shader.has_txq_cube_array_z_comp)
- r600_setup_txq_cube_array_constants(rctx, PIPE_SHADER_GEOMETRY);
+ if (rctx->gs_shader) {
+ need_buf_const = rctx->gs_shader->current->shader.uses_tex_buffers || rctx->gs_shader->current->shader.has_txq_cube_array_z_comp;
+ if (need_buf_const) {
+ if (rctx->b.chip_class < EVERGREEN)
+ r600_setup_buffer_constants(rctx, PIPE_SHADER_GEOMETRY);
+ else
+ eg_setup_buffer_constants(rctx, PIPE_SHADER_GEOMETRY);
+ }
+ }
if (rctx->b.chip_class < EVERGREEN && rctx->ps_shader && rctx->vs_shader) {
if (!r600_adjust_gprs(rctx)) {