v2: Use chip_class instead of family.
v3: Check kernel version for SI.
v4: Preemptively allow amdgpu winsys for SI.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
GL_ARB_arrays_of_arrays DONE (all drivers that support GLSL 1.30)
GL_ARB_ES3_compatibility DONE (all drivers that support GLSL 3.30)
GL_ARB_clear_buffer_object DONE (all drivers)
- GL_ARB_compute_shader DONE (i965)
+ GL_ARB_compute_shader DONE (i965, radeonsi)
GL_ARB_copy_image DONE (i965, nv50, nvc0, r600, radeonsi)
GL_KHR_debug DONE (all drivers)
GL_ARB_explicit_uniform_location DONE (all drivers that support GLSL)
These are the extensions cherry-picked to make GLES 3.1
GLES3.1, GLSL ES 3.1
GL_ARB_arrays_of_arrays DONE (all drivers that support GLSL 1.30)
- GL_ARB_compute_shader DONE (i965)
+ GL_ARB_compute_shader DONE (i965, radeonsi)
GL_ARB_draw_indirect DONE (i965, nvc0, r600, radeonsi, llvmpipe, softpipe)
GL_ARB_explicit_uniform_location DONE (all drivers that support GLSL)
GL_ARB_framebuffer_no_attachments DONE (i965, nvc0, r600, radeonsi, softpipe)
<ul>
<li>OpenGL 4.2 on radeonsi</li>
+<li>GL_ARB_compute_shader on radeonsi</li>
<li>GL_ARB_framebuffer_no_attachments on nvc0, r600, radeonsi, softpipe</li>
<li>GL_ARB_internalformat_query2 on all drivers</li>
<li>GL_ARB_robust_buffer_access_behavior on radeonsi</li>
uint64_t *grid_size = ret;
grid_size[0] = 65535;
grid_size[1] = 65535;
- grid_size[2] = 1;
+ grid_size[2] = 65535;
}
return 3 * sizeof(uint64_t) ;
case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
if (ret) {
uint64_t *block_size = ret;
- block_size[0] = 256;
- block_size[1] = 256;
- block_size[2] = 256;
+ if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
+ ir_type == PIPE_SHADER_IR_TGSI) {
+ block_size[0] = 2048;
+ block_size[1] = 2048;
+ block_size[2] = 2048;
+ } else {
+ block_size[0] = 256;
+ block_size[1] = 256;
+ block_size[2] = 256;
+ }
}
return 3 * sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
if (ret) {
uint64_t *max_threads_per_block = ret;
- *max_threads_per_block = 256;
+ if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
+ ir_type == PIPE_SHADER_IR_TGSI)
+ *max_threads_per_block = 2048;
+ else
+ *max_threads_per_block = 256;
}
return sizeof(uint64_t);
static int si_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enum pipe_shader_cap param)
{
+ struct si_screen *sscreen = (struct si_screen *)pscreen;
+
switch(shader)
{
case PIPE_SHADER_FRAGMENT:
case PIPE_SHADER_CAP_PREFERRED_IR:
return PIPE_SHADER_IR_NATIVE;
- case PIPE_SHADER_CAP_SUPPORTED_IRS:
- return 0;
+ case PIPE_SHADER_CAP_SUPPORTED_IRS: {
+ int ir = 1 << PIPE_SHADER_IR_NATIVE;
+ /* Old kernels disallowed some register writes for SI
+ * that are used for indirect dispatches. */
+ if (HAVE_LLVM >= 0x309 && (sscreen->b.chip_class >= CIK ||
+ sscreen->b.info.drm_major == 3 ||
+ (sscreen->b.info.drm_major == 2 &&
+ sscreen->b.info.drm_minor >= 45)))
+ ir |= 1 << PIPE_SHADER_IR_TGSI;
+
+ return ir;
+ }
case PIPE_SHADER_CAP_DOUBLES:
return HAVE_LLVM >= 0x0307;