#include "program/prog_parameter.h"
#include "program/prog_statevars.h"
#include "intel_batchbuffer.h"
+#include "glsl/glsl_parser_extras.h"
-static void
-upload_vs_state(struct brw_context *brw)
+/**
+ * Creates a streamed BO containing the push constants for the VS or GS on
+ * gen6+.
+ *
+ * Push constants are constant values (such as GLSL uniforms) that are
+ * pre-loaded into a shader stage's register space at thread spawn time.
+ *
+ * Not all GLSL uniforms will be uploaded as push constants: The hardware has
+ * a limitation of 32 or 64 EU registers (256 or 512 floats) per stage to be
+ * uploaded as push constants, while GL 4.4 requires at least 1024 components
+ * to be usable for the VS. Plus, currently we always use pull constants
+ * instead of push constants when doing variable-index array access.
+ *
+ * See brw_curbe.c for the equivalent gen4/5 code.
+ */
+void
+gen6_upload_push_constants(struct brw_context *brw,
+ const struct gl_program *prog,
+ const struct brw_stage_prog_data *prog_data,
+ struct brw_stage_state *stage_state,
+ enum aub_state_struct_type type)
{
- struct intel_context *intel = &brw->intel;
- struct gl_context *ctx = &intel->ctx;
- const struct brw_vertex_program *vp =
- brw_vertex_program_const(brw->vertex_program);
- unsigned int nr_params = brw->vs.prog_data->nr_params / 4;
- drm_intel_bo *constant_bo;
- int i;
+ struct gl_context *ctx = &brw->ctx;
- if (brw->vs.prog_data->nr_params == 0 && !ctx->Transform.ClipPlanesEnabled) {
- /* Disable the push constant buffers. */
- BEGIN_BATCH(5);
- OUT_BATCH(_3DSTATE_CONSTANT_VS << 16 | (5 - 2));
- OUT_BATCH(0);
- OUT_BATCH(0);
- OUT_BATCH(0);
- OUT_BATCH(0);
- ADVANCE_BATCH();
+ if (prog_data->nr_params == 0) {
+ stage_state->push_const_size = 0;
} else {
- int params_uploaded = 0, param_regs;
- float *param;
-
- if (brw->vertex_program->IsNVProgram)
- _mesa_load_tracked_matrices(ctx);
-
/* Updates the ParamaterValues[i] pointers for all parameters of the
* basic type of PROGRAM_STATE_VAR.
*/
- _mesa_load_state_parameters(ctx, vp->program.Base.Parameters);
+ /* XXX: Should this happen somewhere before to get our state flag set? */
+ _mesa_load_state_parameters(ctx, prog->Parameters);
- constant_bo = drm_intel_bo_alloc(intel->bufmgr, "VS constant_bo",
- (MAX_CLIP_PLANES + nr_params) *
- 4 * sizeof(float),
- 4096);
- drm_intel_gem_bo_map_gtt(constant_bo);
- param = constant_bo->virtual;
+ gl_constant_value *param;
+ unsigned i;
- /* This should be loaded like any other param, but it's ad-hoc
- * until we redo the VS backend.
- */
- for (i = 0; i < MAX_CLIP_PLANES; i++) {
- if (ctx->Transform.ClipPlanesEnabled & (1 << i)) {
- memcpy(param, ctx->Transform._ClipUserPlane[i], 4 * sizeof(float));
- param += 4;
- params_uploaded++;
- }
- }
- /* Align to a reg for convenience for brw_vs_emit.c */
- if (params_uploaded & 1) {
- param += 4;
- params_uploaded++;
- }
+ param = brw_state_batch(brw, type,
+ prog_data->nr_params * sizeof(gl_constant_value),
+ 32, &stage_state->push_const_offset);
- for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) {
- if (brw->vs.constant_map[i] != -1) {
- memcpy(param + brw->vs.constant_map[i] * 4,
- vp->program.Base.Parameters->ParameterValues[i],
- 4 * sizeof(float));
- params_uploaded++;
- }
+ STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
+
+ /* _NEW_PROGRAM_CONSTANTS
+ *
+ * Also _NEW_TRANSFORM -- we may reference clip planes other than as a
+ * side effect of dereferencing uniforms, so _NEW_PROGRAM_CONSTANTS
+ * wouldn't be set for them.
+ */
+ for (i = 0; i < prog_data->nr_params; i++) {
+ param[i] = *prog_data->param[i];
}
if (0) {
- printf("VS constant buffer:\n");
- for (i = 0; i < params_uploaded; i++) {
- float *buf = (float *)constant_bo->virtual + i * 4;
- printf("%d: %f %f %f %f\n",
- i, buf[0], buf[1], buf[2], buf[3]);
+ fprintf(stderr, "%s constants:\n",
+ _mesa_shader_stage_to_string(stage_state->stage));
+ for (i = 0; i < prog_data->nr_params; i++) {
+ if ((i & 7) == 0)
+ fprintf(stderr, "g%d: ",
+ prog_data->dispatch_grf_start_reg + i / 8);
+ fprintf(stderr, "%8f ", param[i].f);
+ if ((i & 7) == 7)
+ fprintf(stderr, "\n");
}
+ if ((i & 7) != 0)
+ fprintf(stderr, "\n");
+ fprintf(stderr, "\n");
}
- drm_intel_gem_bo_unmap_gtt(constant_bo);
+ stage_state->push_const_size = ALIGN(prog_data->nr_params, 8) / 8;
+ /* We can only push 32 registers of constants at a time. */
+
+ /* From the SNB PRM (vol2, part 1, section 3.2.1.4: 3DSTATE_CONSTANT_VS:
+ *
+ * "The sum of all four read length fields (each incremented to
+ * represent the actual read length) must be less than or equal to
+ * 32"
+ *
+ * From the IVB PRM (vol2, part 1, section 3.2.1.3: 3DSTATE_CONSTANT_VS:
+ *
+ * "The sum of all four read length fields must be less than or
+ * equal to the size of 64"
+ *
+ * The other shader stages all match the VS's limits.
+ */
+ assert(stage_state->push_const_size <= 32);
+ }
+}
+
+static void
+gen6_upload_vs_push_constants(struct brw_context *brw)
+{
+ struct brw_stage_state *stage_state = &brw->vs.base;
+
+ /* _BRW_NEW_VERTEX_PROGRAM */
+ const struct brw_vertex_program *vp =
+ brw_vertex_program_const(brw->vertex_program);
+ /* BRW_NEW_VS_PROG_DATA */
+ const struct brw_stage_prog_data *prog_data = &brw->vs.prog_data->base.base;
+
+ gen6_upload_push_constants(brw, &vp->program.Base, prog_data,
+ stage_state, AUB_TRACE_VS_CONSTANTS);
+
+ if (brw->gen >= 7) {
+ if (brw->gen == 7 && !brw->is_haswell && !brw->is_baytrail)
+ gen7_emit_vs_workaround_flush(brw);
- param_regs = (params_uploaded + 1) / 2;
- assert(param_regs <= 32);
+ gen7_upload_constant_state(brw, stage_state, true /* active */,
+ _3DSTATE_CONSTANT_VS);
+ }
+}
+
+const struct brw_tracked_state gen6_vs_push_constants = {
+ .dirty = {
+ .mesa = _NEW_PROGRAM_CONSTANTS |
+ _NEW_TRANSFORM,
+ .brw = BRW_NEW_BATCH |
+ BRW_NEW_PUSH_CONSTANT_ALLOCATION |
+ BRW_NEW_VERTEX_PROGRAM |
+ BRW_NEW_VS_PROG_DATA,
+ },
+ .emit = gen6_upload_vs_push_constants,
+};
+static void
+upload_vs_state(struct brw_context *brw)
+{
+ const struct brw_stage_state *stage_state = &brw->vs.base;
+ uint32_t floating_point_mode = 0;
+
+ /* From the BSpec, 3D Pipeline > Geometry > Vertex Shader > State,
+ * 3DSTATE_VS, Dword 5.0 "VS Function Enable":
+ *
+ * [DevSNB] A pipeline flush must be programmed prior to a 3DSTATE_VS
+ * command that causes the VS Function Enable to toggle. Pipeline
+ * flush can be executed by sending a PIPE_CONTROL command with CS
+ * stall bit set and a post sync operation.
+ *
+ * We've already done such a flush at the start of state upload, so we
+ * don't need to do another one here.
+ */
+
+ if (stage_state->push_const_size == 0) {
+ /* Disable the push constant buffers. */
+ BEGIN_BATCH(5);
+ OUT_BATCH(_3DSTATE_CONSTANT_VS << 16 | (5 - 2));
+ OUT_BATCH(0);
+ OUT_BATCH(0);
+ OUT_BATCH(0);
+ OUT_BATCH(0);
+ ADVANCE_BATCH();
+ } else {
BEGIN_BATCH(5);
OUT_BATCH(_3DSTATE_CONSTANT_VS << 16 |
GEN6_CONSTANT_BUFFER_0_ENABLE |
(5 - 2));
- OUT_RELOC(constant_bo,
- I915_GEM_DOMAIN_RENDER, 0, /* XXX: bad domain */
- param_regs - 1);
+ /* Pointer to the VS constant buffer. Covered by the set of
+ * state flags from gen6_upload_vs_constants
+ */
+ OUT_BATCH(stage_state->push_const_offset +
+ stage_state->push_const_size - 1);
OUT_BATCH(0);
OUT_BATCH(0);
OUT_BATCH(0);
ADVANCE_BATCH();
-
- drm_intel_bo_unreference(constant_bo);
}
+ if (brw->vs.prog_data->base.base.use_alt_mode)
+ floating_point_mode = GEN6_VS_FLOATING_POINT_MODE_ALT;
+
BEGIN_BATCH(6);
OUT_BATCH(_3DSTATE_VS << 16 | (6 - 2));
- OUT_RELOC(brw->vs.prog_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
- OUT_BATCH((0 << GEN6_VS_SAMPLER_COUNT_SHIFT) |
- GEN6_VS_FLOATING_POINT_MODE_ALT |
- (brw->vs.nr_surfaces << GEN6_VS_BINDING_TABLE_ENTRY_COUNT_SHIFT));
- OUT_BATCH(0); /* scratch space base offset */
- OUT_BATCH((1 << GEN6_VS_DISPATCH_START_GRF_SHIFT) |
- (brw->vs.prog_data->urb_read_length << GEN6_VS_URB_READ_LENGTH_SHIFT) |
+ OUT_BATCH(stage_state->prog_offset);
+ OUT_BATCH(floating_point_mode |
+ ((ALIGN(stage_state->sampler_count, 4)/4) << GEN6_VS_SAMPLER_COUNT_SHIFT) |
+ ((brw->vs.prog_data->base.base.binding_table.size_bytes / 4) <<
+ GEN6_VS_BINDING_TABLE_ENTRY_COUNT_SHIFT));
+
+ if (brw->vs.prog_data->base.base.total_scratch) {
+ OUT_RELOC(stage_state->scratch_bo,
+ I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
+ ffs(brw->vs.prog_data->base.base.total_scratch) - 11);
+ } else {
+ OUT_BATCH(0);
+ }
+
+ OUT_BATCH((brw->vs.prog_data->base.base.dispatch_grf_start_reg <<
+ GEN6_VS_DISPATCH_START_GRF_SHIFT) |
+ (brw->vs.prog_data->base.urb_read_length << GEN6_VS_URB_READ_LENGTH_SHIFT) |
(0 << GEN6_VS_URB_ENTRY_READ_OFFSET_SHIFT));
- OUT_BATCH((0 << GEN6_VS_MAX_THREADS_SHIFT) |
+
+ OUT_BATCH(((brw->max_vs_threads - 1) << GEN6_VS_MAX_THREADS_SHIFT) |
GEN6_VS_STATISTICS_ENABLE |
GEN6_VS_ENABLE);
ADVANCE_BATCH();
+
+ /* Based on my reading of the simulator, the VS constants don't get
+ * pulled into the VS FF unit until an appropriate pipeline flush
+ * happens, and instead the 3DSTATE_CONSTANT_VS packet just adds
+ * references to them into a little FIFO. The flushes are common,
+ * but don't reliably happen between this and a 3DPRIMITIVE, causing
+ * the primitive to use the wrong constants. Then the FIFO
+ * containing the constant setup gets added to again on the next
+ * constants change, and eventually when a flush does happen the
+ * unit is overwhelmed by constant changes and dies.
+ *
+ * To avoid this, send a PIPE_CONTROL down the line that will
+ * update the unit immediately loading the constants. The flush
+ * type bits here were those set by the STATE_BASE_ADDRESS whose
+ * move in a82a43e8d99e1715dd11c9c091b5ab734079b6a6 triggered the
+ * bug reports that led to this workaround, and may be more than
+ * what is strictly required to avoid the issue.
+ */
+ brw_emit_pipe_control_flush(brw,
+ PIPE_CONTROL_DEPTH_STALL |
+ PIPE_CONTROL_INSTRUCTION_INVALIDATE |
+ PIPE_CONTROL_STATE_CACHE_INVALIDATE);
}
const struct brw_tracked_state gen6_vs_state = {
.dirty = {
- .mesa = _NEW_TRANSFORM | _NEW_PROGRAM_CONSTANTS,
- .brw = (BRW_NEW_CURBE_OFFSETS |
- BRW_NEW_NR_VS_SURFACES |
- BRW_NEW_URB_FENCE |
- BRW_NEW_CONTEXT),
- .cache = CACHE_NEW_VS_PROG
+ .mesa = _NEW_PROGRAM_CONSTANTS |
+ _NEW_TRANSFORM,
+ .brw = BRW_NEW_BATCH |
+ BRW_NEW_CONTEXT |
+ BRW_NEW_PUSH_CONSTANT_ALLOCATION |
+ BRW_NEW_VERTEX_PROGRAM |
+ BRW_NEW_VS_PROG_DATA,
},
.emit = upload_vs_state,
};