* Keith Whitwell <keith@tungstengraphics.com>
*/
+#include <pthread.h>
#include "main/imports.h"
#include "main/enums.h"
#include "main/shaderobj.h"
#include "brw_context.h"
#include "brw_wm.h"
+static unsigned
+get_new_program_id(struct intel_screen *screen)
+{
+ static pthread_mutex_t m = PTHREAD_MUTEX_INITIALIZER;
+ pthread_mutex_lock(&m);
+ unsigned id = screen->program_id++;
+ pthread_mutex_unlock(&m);
+ return id;
+}
+
static void brwBindProgram( struct gl_context *ctx,
GLenum target,
struct gl_program *prog )
case GL_VERTEX_PROGRAM_ARB:
brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
break;
+ case MESA_GEOMETRY_PROGRAM:
+ brw->state.dirty.brw |= BRW_NEW_GEOMETRY_PROGRAM;
+ break;
case GL_FRAGMENT_PROGRAM_ARB:
brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
break;
case GL_VERTEX_PROGRAM_ARB: {
struct brw_vertex_program *prog = CALLOC_STRUCT(brw_vertex_program);
if (prog) {
- prog->id = brw->program_id++;
+ prog->id = get_new_program_id(brw->intelScreen);
return _mesa_init_vertex_program( ctx, &prog->program,
target, id );
case GL_FRAGMENT_PROGRAM_ARB: {
struct brw_fragment_program *prog = CALLOC_STRUCT(brw_fragment_program);
if (prog) {
- prog->id = brw->program_id++;
+ prog->id = get_new_program_id(brw->intelScreen);
return _mesa_init_fragment_program( ctx, &prog->program,
target, id );
return NULL;
}
+ case MESA_GEOMETRY_PROGRAM: {
+ struct brw_geometry_program *prog = CALLOC_STRUCT(brw_geometry_program);
+ if (prog) {
+ prog->id = get_new_program_id(brw->intelScreen);
+
+ return _mesa_init_geometry_program(ctx, &prog->program, target, id);
+ } else {
+ return NULL;
+ }
+ }
+
default:
- return _mesa_new_program(ctx, target, id);
+ assert(!"Unsupported target in brwNewProgram()");
+ return NULL;
}
}
{
struct brw_context *brw = brw_context(ctx);
- if (target == GL_FRAGMENT_PROGRAM_ARB) {
+ switch (target) {
+ case GL_FRAGMENT_PROGRAM_ARB: {
struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
struct brw_fragment_program *newFP = brw_fragment_program(fprog);
const struct brw_fragment_program *curFP =
if (newFP == curFP)
brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
- newFP->id = brw->program_id++;
+ newFP->id = get_new_program_id(brw->intelScreen);
+ break;
}
- else if (target == GL_VERTEX_PROGRAM_ARB) {
+ case GL_VERTEX_PROGRAM_ARB: {
struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
struct brw_vertex_program *newVP = brw_vertex_program(vprog);
const struct brw_vertex_program *curVP =
if (newVP->program.IsPositionInvariant) {
_mesa_insert_mvp_code(ctx, &newVP->program);
}
- newVP->id = brw->program_id++;
+ newVP->id = get_new_program_id(brw->intelScreen);
/* Also tell tnl about it:
*/
_tnl_program_string(ctx, target, prog);
+ break;
}
+ default:
+ /*
+ * driver->ProgramStringNotify is only called for ARB programs, fixed
+ * function vertex programs, and ir_to_mesa (which isn't used by the
+ * i965 back-end). Therefore, even after geometry shaders are added,
+ * this function should only ever be called with a target of
+ * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
+ */
+ assert(!"Unexpected target in brwProgramStringNotify");
+ break;
+ }
+
+ brw_add_texrect_params(prog);
return true;
}
+void
+brw_add_texrect_params(struct gl_program *prog)
+{
+ for (int texunit = 0; texunit < BRW_MAX_TEX_UNIT; texunit++) {
+ if (!(prog->TexturesUsed[texunit] & (1 << TEXTURE_RECT_INDEX)))
+ continue;
+
+ int tokens[STATE_LENGTH] = {
+ STATE_INTERNAL,
+ STATE_TEXRECT_SCALE,
+ texunit,
+ 0,
+ 0
+ };
+
+ _mesa_add_state_reference(prog->Parameters, (gl_state_index *)tokens);
+ }
+}
+
/* Per-thread scratch space is a power-of-two multiple of 1KB. */
int
brw_get_scratch_size(int size)
}
void
-brw_get_scratch_bo(struct intel_context *intel,
+brw_get_scratch_bo(struct brw_context *brw,
drm_intel_bo **scratch_bo, int size)
{
drm_intel_bo *old_bo = *scratch_bo;
}
if (!old_bo) {
- *scratch_bo = drm_intel_bo_alloc(intel->bufmgr, "scratch bo", size, 4096);
+ *scratch_bo = drm_intel_bo_alloc(brw->bufmgr, "scratch bo", size, 4096);
}
}
void
brw_init_shader_time(struct brw_context *brw)
{
- struct intel_context *intel = &brw->intel;
-
const int max_entries = 4096;
- brw->shader_time.bo = drm_intel_bo_alloc(intel->bufmgr, "shader time",
- max_entries * 4, 4096);
- brw->shader_time.programs = rzalloc_array(brw, struct gl_shader_program *,
+ brw->shader_time.bo = drm_intel_bo_alloc(brw->bufmgr, "shader time",
+ max_entries * SHADER_TIME_STRIDE,
+ 4096);
+ brw->shader_time.shader_programs = rzalloc_array(brw, struct gl_shader_program *,
+ max_entries);
+ brw->shader_time.programs = rzalloc_array(brw, struct gl_program *,
max_entries);
brw->shader_time.types = rzalloc_array(brw, enum shader_time_shader_type,
max_entries);
return 1;
}
+static void
+get_written_and_reset(struct brw_context *brw, int i,
+ uint64_t *written, uint64_t *reset)
+{
+ enum shader_time_shader_type type = brw->shader_time.types[i];
+ assert(type == ST_VS || type == ST_FS8 || type == ST_FS16);
+
+ /* Find where we recorded written and reset. */
+ int wi, ri;
+
+ for (wi = i; brw->shader_time.types[wi] != type + 1; wi++)
+ ;
+
+ for (ri = i; brw->shader_time.types[ri] != type + 2; ri++)
+ ;
+
+ *written = brw->shader_time.cumulative[wi];
+ *reset = brw->shader_time.cumulative[ri];
+}
+
+static void
+print_shader_time_line(const char *stage, const char *name,
+ int shader_num, uint64_t time, uint64_t total)
+{
+ printf("%-6s%-6s", stage, name);
+
+ if (shader_num != -1)
+ printf("%4d: ", shader_num);
+ else
+ printf(" : ");
+
+ printf("%16lld (%7.2f Gcycles) %4.1f%%\n",
+ (long long)time,
+ (double)time / 1000000000.0,
+ (double)time / total * 100.0);
+}
+
static void
brw_report_shader_time(struct brw_context *brw)
{
if (!brw->shader_time.bo || !brw->shader_time.num_entries)
return;
+ uint64_t scaled[brw->shader_time.num_entries];
uint64_t *sorted[brw->shader_time.num_entries];
+ uint64_t total_by_type[ST_FS16 + 1];
+ memset(total_by_type, 0, sizeof(total_by_type));
double total = 0;
for (int i = 0; i < brw->shader_time.num_entries; i++) {
- sorted[i] = &brw->shader_time.cumulative[i];
- total += brw->shader_time.cumulative[i];
+ uint64_t written = 0, reset = 0;
+ enum shader_time_shader_type type = brw->shader_time.types[i];
+
+ sorted[i] = &scaled[i];
+
+ switch (type) {
+ case ST_VS_WRITTEN:
+ case ST_VS_RESET:
+ case ST_FS8_WRITTEN:
+ case ST_FS8_RESET:
+ case ST_FS16_WRITTEN:
+ case ST_FS16_RESET:
+ /* We'll handle these when along with the time. */
+ scaled[i] = 0;
+ continue;
+
+ case ST_VS:
+ case ST_FS8:
+ case ST_FS16:
+ get_written_and_reset(brw, i, &written, &reset);
+ break;
+
+ default:
+ /* I sometimes want to print things that aren't the 3 shader times.
+ * Just print the sum in that case.
+ */
+ written = 1;
+ reset = 0;
+ break;
+ }
+
+ uint64_t time = brw->shader_time.cumulative[i];
+ if (written) {
+ scaled[i] = time / written * (written + reset);
+ } else {
+ scaled[i] = time;
+ }
+
+ switch (type) {
+ case ST_VS:
+ case ST_FS8:
+ case ST_FS16:
+ total_by_type[type] += scaled[i];
+ break;
+ default:
+ break;
+ }
+
+ total += scaled[i];
}
if (total == 0) {
qsort(sorted, brw->shader_time.num_entries, sizeof(sorted[0]), compare_time);
printf("\n");
- printf("type ID cycles spent %% of total\n");
+ printf("type ID cycles spent %% of total\n");
for (int s = 0; s < brw->shader_time.num_entries; s++) {
+ const char *shader_name;
+ const char *stage;
/* Work back from the sorted pointers times to a time to print. */
- int i = sorted[s] - brw->shader_time.cumulative;
+ int i = sorted[s] - scaled;
+
+ if (scaled[i] == 0)
+ continue;
int shader_num = -1;
- if (brw->shader_time.programs[i]) {
- shader_num = brw->shader_time.programs[i]->Name;
+ if (brw->shader_time.shader_programs[i]) {
+ shader_num = brw->shader_time.shader_programs[i]->Name;
+
+ /* The fixed function fragment shader generates GLSL IR with a Name
+ * of 0, and nothing else does.
+ */
+ if (shader_num == 0 &&
+ (brw->shader_time.types[i] == ST_FS8 ||
+ brw->shader_time.types[i] == ST_FS16)) {
+ shader_name = "ff";
+ shader_num = -1;
+ } else {
+ shader_name = "glsl";
+ }
+ } else if (brw->shader_time.programs[i]) {
+ shader_num = brw->shader_time.programs[i]->Id;
+ if (shader_num == 0) {
+ shader_name = "ff";
+ shader_num = -1;
+ } else {
+ shader_name = "prog";
+ }
+ } else {
+ shader_name = "other";
}
switch (brw->shader_time.types[i]) {
case ST_VS:
- printf("vs %4d: ", shader_num);
+ stage = "vs";
break;
case ST_FS8:
- printf("fs8 %4d: ", shader_num);
+ stage = "fs8";
break;
case ST_FS16:
- printf("fs16 %4d: ", shader_num);
+ stage = "fs16";
break;
default:
- printf("other: ");
+ stage = "other";
break;
}
- printf("%16lld (%7.2f Gcycles) %4.1f%%\n",
- (long long)brw->shader_time.cumulative[i],
- (double)brw->shader_time.cumulative[i] / 1000000000.0,
- (double)brw->shader_time.cumulative[i] / total * 100.0);
+ print_shader_time_line(stage, shader_name, shader_num,
+ scaled[i], total);
}
+
+ printf("\n");
+ print_shader_time_line("total", "vs", -1, total_by_type[ST_VS], total);
+ print_shader_time_line("total", "fs8", -1, total_by_type[ST_FS8], total);
+ print_shader_time_line("total", "fs16", -1, total_by_type[ST_FS16], total);
}
static void
uint32_t *times = brw->shader_time.bo->virtual;
for (int i = 0; i < brw->shader_time.num_entries; i++) {
- brw->shader_time.cumulative[i] += times[i];
+ brw->shader_time.cumulative[i] += times[i * SHADER_TIME_STRIDE / 4];
}
/* Zero the BO out to clear it out for our next collection.
}
}
+/**
+ * Chooses an index in the shader_time buffer and sets up tracking information
+ * for our printouts.
+ *
+ * Note that this holds on to references to the underlying programs, which may
+ * change their lifetimes compared to normal operation.
+ */
+int
+brw_get_shader_time_index(struct brw_context *brw,
+ struct gl_shader_program *shader_prog,
+ struct gl_program *prog,
+ enum shader_time_shader_type type)
+{
+ struct gl_context *ctx = &brw->ctx;
+
+ int shader_time_index = brw->shader_time.num_entries++;
+ assert(shader_time_index < brw->shader_time.max_entries);
+ brw->shader_time.types[shader_time_index] = type;
+
+ _mesa_reference_shader_program(ctx,
+ &brw->shader_time.shader_programs[shader_time_index],
+ shader_prog);
+
+ _mesa_reference_program(ctx,
+ &brw->shader_time.programs[shader_time_index],
+ prog);
+
+ return shader_time_index;
+}
+
void
brw_destroy_shader_time(struct brw_context *brw)
{
projects / mesa.git / blobdiff