#include "spu_tile.h"
#include "spu_vertex_shader.h"
#include "spu_dcache.h"
-#include "spu_debug.h"
#include "cell/common.h"
+static INLINE int
+align(int value, int alignment)
+{
+ return (value + alignment - 1) & ~(alignment - 1);
+}
+
+
+
/**
* Tell the PPU that this SPU has finished copying a buffer to
* local store and that it may be reused by the PPU.
release_buffer(uint buffer)
{
/* Evidently, using less than a 16-byte status doesn't work reliably */
- static const uint status[4] ALIGN16_ATTRIB
- = {CELL_BUFFER_STATUS_FREE, 0, 0, 0};
-
+ static const vector unsigned int status = {CELL_BUFFER_STATUS_FREE,
+ CELL_BUFFER_STATUS_FREE,
+ CELL_BUFFER_STATUS_FREE,
+ CELL_BUFFER_STATUS_FREE};
const uint index = 4 * (spu.init.id * CELL_NUM_BUFFERS + buffer);
uint *dst = spu.init.buffer_status + index;
}
+/**
+ * Write CELL_FENCE_SIGNALLED back to the fence status qword in main memory.
+ * There's a qword of status per SPU.
+ */
+static void
+cmd_fence(struct cell_command_fence *fence_cmd)
+{
+ static const vector unsigned int status = {CELL_FENCE_SIGNALLED,
+ CELL_FENCE_SIGNALLED,
+ CELL_FENCE_SIGNALLED,
+ CELL_FENCE_SIGNALLED};
+ uint *dst = (uint *) fence_cmd->fence;
+ dst += 4 * spu.init.id; /* main store/memory address, not local store */
+ ASSERT_ALIGN16(dst);
+ mfc_put((void *) &status, /* src in local memory */
+ (unsigned int) dst, /* dst in main memory */
+ sizeof(status), /* size */
+ TAG_FENCE, /* tag */
+ 0, /* tid */
+ 0 /* rid */);
+}
+
+
static void
cmd_clear_surface(const struct cell_command_clear_surface *clear)
{
- DEBUG_PRINTF("CLEAR SURF %u to 0x%08x\n", clear->surface, clear->value);
+ D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF %u to 0x%08x\n", clear->surface, clear->value);
if (clear->surface == 0) {
spu.fb.color_clear_value = clear->value;
#endif /* CLEAR_OPT */
- DEBUG_PRINTF("CLEAR SURF done\n");
+ D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF done\n");
}
static void
cmd_release_verts(const struct cell_command_release_verts *release)
{
- DEBUG_PRINTF("RELEASE VERTS %u\n", release->vertex_buf);
+ D_PRINTF(CELL_DEBUG_CMD, "RELEASE VERTS %u\n", release->vertex_buf);
ASSERT(release->vertex_buf != ~0U);
release_buffer(release->vertex_buf);
}
{
static int warned = 0;
- DEBUG_PRINTF("CMD_STATE_FRAGMENT_OPS\n");
+ D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_OPS\n");
/* Copy SPU code from batch buffer to spu buffer */
- memcpy(spu.fragment_ops_code, fops->code, SPU_MAX_FRAGMENT_OPS_INSTS * 4);
+ memcpy(spu.fragment_ops_code_front, fops->code_front, SPU_MAX_FRAGMENT_OPS_INSTS * 4);
+ memcpy(spu.fragment_ops_code_back, fops->code_back, SPU_MAX_FRAGMENT_OPS_INSTS * 4);
/* Copy state info (for fallback case only) */
memcpy(&spu.depth_stencil_alpha, &fops->dsa, sizeof(fops->dsa));
memcpy(&spu.blend, &fops->blend, sizeof(fops->blend));
+ memcpy(&spu.blend_color, &fops->blend_color, sizeof(fops->blend_color));
/* Parity twist! For now, always use the fallback code by default,
* only switching to codegen when specifically requested. This
* raw state records that the fallback code requires.
*/
if ((spu.init.debug_flags & CELL_DEBUG_FRAGMENT_OP_FALLBACK) == 0) {
- spu.fragment_ops = (spu_fragment_ops_func) spu.fragment_ops_code;
+ spu.fragment_ops[CELL_FACING_FRONT] = (spu_fragment_ops_func) spu.fragment_ops_code_front;
+ spu.fragment_ops[CELL_FACING_BACK] = (spu_fragment_ops_func) spu.fragment_ops_code_back;
}
else {
/* otherwise, the default fallback code remains in place */
}
}
- spu.read_depth = spu.depth_stencil_alpha.depth.enabled;
- spu.read_stencil = spu.depth_stencil_alpha.stencil[0].enabled;
+ spu.read_depth_stencil = (spu.depth_stencil_alpha.depth.enabled || spu.depth_stencil_alpha.stencil[0].enabled);
}
static void
cmd_state_fragment_program(const struct cell_command_fragment_program *fp)
{
- DEBUG_PRINTF("CMD_STATE_FRAGMENT_PROGRAM\n");
+ D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_PROGRAM\n");
/* Copy SPU code from batch buffer to spu buffer */
memcpy(spu.fragment_program_code, fp->code,
SPU_MAX_FRAGMENT_PROGRAM_INSTS * 4);
}
+static uint
+cmd_state_fs_constants(const uint64_t *buffer, uint pos)
+{
+ const uint num_const = buffer[pos + 1];
+ const float *constants = (const float *) &buffer[pos + 2];
+ uint i;
+
+ D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FS_CONSTANTS (%u)\n", num_const);
+
+ /* Expand each float to float[4] for SOA execution */
+ for (i = 0; i < num_const; i++) {
+ D_PRINTF(CELL_DEBUG_CMD, " const[%u] = %f\n", i, constants[i]);
+ spu.constants[i] = spu_splats(constants[i]);
+ }
+
+ /* return new buffer pos (in 8-byte words) */
+ return pos + 2 + num_const / 2;
+}
+
+
static void
cmd_state_framebuffer(const struct cell_command_framebuffer *cmd)
{
- DEBUG_PRINTF("FRAMEBUFFER: %d x %d at %p, cformat 0x%x zformat 0x%x\n",
+ D_PRINTF(CELL_DEBUG_CMD, "FRAMEBUFFER: %d x %d at %p, cformat 0x%x zformat 0x%x\n",
cmd->width,
cmd->height,
cmd->color_start,
}
+/**
+ * Tex texture mask_s/t and scale_s/t fields depend on the texture size and
+ * sampler wrap modes.
+ */
+static void
+update_tex_masks(struct spu_texture *texture,
+ const struct pipe_sampler_state *sampler)
+{
+ uint i;
+
+ for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) {
+ int width = texture->level[i].width;
+ int height = texture->level[i].height;
+
+ if (sampler->wrap_s == PIPE_TEX_WRAP_REPEAT)
+ texture->level[i].mask_s = spu_splats(width - 1);
+ else
+ texture->level[i].mask_s = spu_splats(~0);
+
+ if (sampler->wrap_t == PIPE_TEX_WRAP_REPEAT)
+ texture->level[i].mask_t = spu_splats(height - 1);
+ else
+ texture->level[i].mask_t = spu_splats(~0);
+
+ if (sampler->normalized_coords) {
+ texture->level[i].scale_s = spu_splats((float) width);
+ texture->level[i].scale_t = spu_splats((float) height);
+ }
+ else {
+ texture->level[i].scale_s = spu_splats(1.0f);
+ texture->level[i].scale_t = spu_splats(1.0f);
+ }
+ }
+}
+
+
static void
cmd_state_sampler(const struct cell_command_sampler *sampler)
{
- DEBUG_PRINTF("SAMPLER [%u]\n", sampler->unit);
+ uint unit = sampler->unit;
- spu.sampler[sampler->unit] = sampler->state;
- if (spu.sampler[sampler->unit].min_img_filter == PIPE_TEX_FILTER_LINEAR)
- spu.sample_texture[sampler->unit] = sample_texture_bilinear;
- else
- spu.sample_texture[sampler->unit] = sample_texture_nearest;
+ D_PRINTF(CELL_DEBUG_CMD, "SAMPLER [%u]\n", unit);
+
+ spu.sampler[unit] = sampler->state;
+
+ switch (spu.sampler[unit].min_img_filter) {
+ case PIPE_TEX_FILTER_LINEAR:
+ spu.min_sample_texture_2d[unit] = sample_texture_2d_bilinear;
+ break;
+ case PIPE_TEX_FILTER_ANISO:
+ /* fall-through, for now */
+ case PIPE_TEX_FILTER_NEAREST:
+ spu.min_sample_texture_2d[unit] = sample_texture_2d_nearest;
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ switch (spu.sampler[sampler->unit].mag_img_filter) {
+ case PIPE_TEX_FILTER_LINEAR:
+ spu.mag_sample_texture_2d[unit] = sample_texture_2d_bilinear;
+ break;
+ case PIPE_TEX_FILTER_ANISO:
+ /* fall-through, for now */
+ case PIPE_TEX_FILTER_NEAREST:
+ spu.mag_sample_texture_2d[unit] = sample_texture_2d_nearest;
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ switch (spu.sampler[sampler->unit].min_mip_filter) {
+ case PIPE_TEX_MIPFILTER_NEAREST:
+ case PIPE_TEX_MIPFILTER_LINEAR:
+ spu.sample_texture_2d[unit] = sample_texture_2d_lod;
+ break;
+ case PIPE_TEX_MIPFILTER_NONE:
+ spu.sample_texture_2d[unit] = spu.mag_sample_texture_2d[unit];
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ update_tex_masks(&spu.texture[unit], &spu.sampler[unit]);
}
cmd_state_texture(const struct cell_command_texture *texture)
{
const uint unit = texture->unit;
- const uint width = texture->width;
- const uint height = texture->height;
+ uint i;
+
+ D_PRINTF(CELL_DEBUG_CMD, "TEXTURE [%u]\n", texture->unit);
+
+ spu.texture[unit].max_level = 0;
+ spu.texture[unit].target = texture->target;
+
+ for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) {
+ uint width = texture->width[i];
+ uint height = texture->height[i];
+ uint depth = texture->depth[i];
- DEBUG_PRINTF("TEXTURE [%u] at %p size %u x %u\n",
- texture->unit, texture->start,
- texture->width, texture->height);
+ D_PRINTF(CELL_DEBUG_CMD, " LEVEL %u: at %p size[0] %u x %u\n", i,
+ texture->start[i], texture->width[i], texture->height[i]);
- spu.texture[unit].start = texture->start;
- spu.texture[unit].width = width;
- spu.texture[unit].height = height;
+ spu.texture[unit].level[i].start = texture->start[i];
+ spu.texture[unit].level[i].width = width;
+ spu.texture[unit].level[i].height = height;
+ spu.texture[unit].level[i].depth = depth;
- spu.texture[unit].tiles_per_row = width / TILE_SIZE;
+ spu.texture[unit].level[i].tiles_per_row =
+ (width + TILE_SIZE - 1) / TILE_SIZE;
- spu.texture[unit].tex_size = (vector float) { width, height, 0.0, 0.0};
- spu.texture[unit].tex_size_mask = (vector unsigned int)
- { width - 1, height - 1, 0, 0 };
- spu.texture[unit].tex_size_x_mask = spu_splats(width - 1);
- spu.texture[unit].tex_size_y_mask = spu_splats(height - 1);
+ spu.texture[unit].level[i].bytes_per_image =
+ 4 * align(width, TILE_SIZE) * align(height, TILE_SIZE) * depth;
+
+ spu.texture[unit].level[i].max_s = spu_splats((int) width - 1);
+ spu.texture[unit].level[i].max_t = spu_splats((int) height - 1);
+
+ if (texture->start[i])
+ spu.texture[unit].max_level = i;
+ }
+
+ update_tex_masks(&spu.texture[unit], &spu.sampler[unit]);
}
static void
cmd_state_vertex_info(const struct vertex_info *vinfo)
{
- DEBUG_PRINTF("VERTEX_INFO num_attribs=%u\n", vinfo->num_attribs);
+ D_PRINTF(CELL_DEBUG_CMD, "VERTEX_INFO num_attribs=%u\n", vinfo->num_attribs);
ASSERT(vinfo->num_attribs >= 1);
ASSERT(vinfo->num_attribs <= 8);
memcpy(&spu.vertex_info, vinfo, sizeof(*vinfo));
static void
cmd_finish(void)
{
- DEBUG_PRINTF("FINISH\n");
+ D_PRINTF(CELL_DEBUG_CMD, "FINISH\n");
really_clear_tiles(0);
/* wait for all outstanding DMAs to finish */
mfc_write_tag_mask(~0);
const unsigned usize = size / sizeof(buffer[0]);
uint pos;
- DEBUG_PRINTF("BATCH buffer %u, len %u, from %p\n",
+ D_PRINTF(CELL_DEBUG_CMD, "BATCH buffer %u, len %u, from %p\n",
buf, size, spu.init.buffers[buf]);
ASSERT((opcode & CELL_CMD_OPCODE_MASK) == CELL_CMD_BATCH);
wait_on_mask(1 << TAG_BATCH_BUFFER);
/* Tell PPU we're done copying the buffer to local store */
- DEBUG_PRINTF("release batch buf %u\n", buf);
+ D_PRINTF(CELL_DEBUG_CMD, "release batch buf %u\n", buf);
release_buffer(buf);
/*
pos += sizeof(*fp) / 8;
}
break;
+ case CELL_CMD_STATE_FS_CONSTANTS:
+ pos = cmd_state_fs_constants(buffer, pos);
+ break;
+ case CELL_CMD_STATE_RASTERIZER:
+ {
+ struct cell_command_rasterizer *rast =
+ (struct cell_command_rasterizer *) &buffer[pos];
+ spu.rasterizer = rast->rasterizer;
+ pos += sizeof(*rast) / 8;
+ }
+ break;
case CELL_CMD_STATE_SAMPLER:
{
struct cell_command_sampler *sampler
cmd_finish();
pos += 1;
break;
+ case CELL_CMD_FENCE:
+ {
+ struct cell_command_fence *fence_cmd =
+ (struct cell_command_fence *) &buffer[pos];
+ cmd_fence(fence_cmd);
+ pos += sizeof(*fence_cmd) / 8;
+ }
+ break;
case CELL_CMD_RELEASE_VERTS:
{
struct cell_command_release_verts *release
}
}
- DEBUG_PRINTF("BATCH complete\n");
+ D_PRINTF(CELL_DEBUG_CMD, "BATCH complete\n");
}
+#define PERF 0
+
/**
* Main loop for SPEs: Get a command, execute it, repeat.
void
command_loop(void)
{
- struct cell_command cmd;
int exitFlag = 0;
+ uint t0, t1;
- DEBUG_PRINTF("Enter command loop\n");
-
- ASSERT((sizeof(struct cell_command) & 0xf) == 0);
- ASSERT_ALIGN16(&cmd);
+ D_PRINTF(CELL_DEBUG_CMD, "Enter command loop\n");
while (!exitFlag) {
unsigned opcode;
- int tag = 0;
- DEBUG_PRINTF("Wait for cmd...\n");
+ D_PRINTF(CELL_DEBUG_CMD, "Wait for cmd...\n");
+
+ if (PERF)
+ spu_write_decrementer(~0);
/* read/wait from mailbox */
opcode = (unsigned int) spu_read_in_mbox();
+ D_PRINTF(CELL_DEBUG_CMD, "got cmd 0x%x\n", opcode);
- DEBUG_PRINTF("got cmd 0x%x\n", opcode);
-
- /* command payload */
- mfc_get(&cmd, /* dest */
- (unsigned int) spu.init.cmd, /* src */
- sizeof(struct cell_command), /* bytes */
- tag,
- 0, /* tid */
- 0 /* rid */);
- wait_on_mask( 1 << tag );
-
- /*
- * NOTE: most commands should be contained in a batch buffer
- */
+ if (PERF)
+ t0 = spu_read_decrementer();
switch (opcode & CELL_CMD_OPCODE_MASK) {
case CELL_CMD_EXIT:
- DEBUG_PRINTF("EXIT\n");
+ D_PRINTF(CELL_DEBUG_CMD, "EXIT\n");
exitFlag = 1;
break;
case CELL_CMD_VS_EXECUTE:
printf("Bad opcode 0x%x!\n", opcode & CELL_CMD_OPCODE_MASK);
}
+ if (PERF) {
+ t1 = spu_read_decrementer();
+ printf("wait mbox time: %gms batch time: %gms\n",
+ (~0u - t0) * spu.init.inv_timebase,
+ (t0 - t1) * spu.init.inv_timebase);
+ }
}
- DEBUG_PRINTF("Exit command loop\n");
+ D_PRINTF(CELL_DEBUG_CMD, "Exit command loop\n");
- spu_dcache_report();
+ if (spu.init.debug_flags & CELL_DEBUG_CACHE)
+ spu_dcache_report();
}
projects / mesa.git / blobdiff