diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index 7da80da6829..29c2180d010 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -239,7 +239,7 @@ static void *si_create_compute_state(struct pipe_context *ctx, const struct pipe
       si_sampler_and_image_descriptors_idx(PIPE_SHADER_COMPUTE);
    sel->info.base.shared_size = cso->req_local_mem;
    program->shader.selector = &program->sel;
-   program->shader.wave_size = si_get_shader_wave_size(&program->shader);
+   program->shader.wave_size = si_get_shader_wave_size(sscreen, &program->shader);
    program->ir_type = cso->ir_type;
    program->private_size = cso->req_private_mem;
    program->input_size = cso->req_input_mem;
diff --git a/src/gallium/drivers/radeonsi/si_compute_blit.c b/src/gallium/drivers/radeonsi/si_compute_blit.c
index 42f293fafca..0ab9993a7e0 100644
--- a/src/gallium/drivers/radeonsi/si_compute_blit.c
+++ b/src/gallium/drivers/radeonsi/si_compute_blit.c
@@ -305,7 +305,8 @@ static void si_compute_do_clear_or_copy(struct si_context *sctx, struct pipe_res
       src ? SI_COMPUTE_COPY_DW_PER_THREAD : SI_COMPUTE_CLEAR_DW_PER_THREAD;
    unsigned instructions_per_thread = MAX2(1, dwords_per_thread / 4);
    unsigned dwords_per_instruction = dwords_per_thread / instructions_per_thread;
-   unsigned block_size = sctx->screen->compute_wave_size;
+   /* The shader declares the block size like this: */
+   unsigned block_size = si_get_shader_wave_size(sctx->screen, NULL);
    unsigned dwords_per_wave = dwords_per_thread * block_size;
 
    unsigned num_dwords = size / 4;
@@ -598,9 +599,10 @@ void si_compute_copy_image(struct si_context *sctx, struct pipe_resource *dst, u
       info.block[1] = ssrc->surface.u.gfx9.color.dcc_block_height;
       info.block[2] = ssrc->surface.u.gfx9.color.dcc_block_depth;
 
+      unsigned default_wave_size = si_get_shader_wave_size(sctx->screen, NULL);;
+
       /* Make sure the block size is at least the same as wave size. */
-      while (info.block[0] * info.block[1] * info.block[2] <
-             sctx->screen->compute_wave_size) {
+      while (info.block[0] * info.block[1] * info.block[2] < default_wave_size) {
          info.block[0] *= 2;
       }
 
diff --git a/src/gallium/drivers/radeonsi/si_get.c b/src/gallium/drivers/radeonsi/si_get.c
index 18726d2cb5d..a931af4b3d4 100644
--- a/src/gallium/drivers/radeonsi/si_get.c
+++ b/src/gallium/drivers/radeonsi/si_get.c
@@ -881,7 +881,7 @@ static int si_get_compute_param(struct pipe_screen *screen, enum pipe_shader_ir
    case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
       if (ret) {
          uint32_t *subgroup_size = ret;
-         *subgroup_size = sscreen->compute_wave_size;
+         *subgroup_size = si_get_shader_wave_size(sscreen, NULL);
       }
       return sizeof(uint32_t);
    case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 793d74c7b2a..f58a48d4b48 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -1995,19 +1995,20 @@ static inline unsigned si_get_wave_size(struct si_screen *sscreen,
       return sscreen->ge_wave_size;
 }
 
-static inline unsigned si_get_shader_wave_size(struct si_shader *shader)
+static inline unsigned si_get_shader_wave_size(struct si_screen *sscreen, struct si_shader *shader)
 {
-   if (shader->is_gs_copy_shader)
+   /* There are a few uses that pass shader=NULL here, expecting the default compute wave size. */
+   struct si_shader_info *info = shader ? &shader->selector->info : NULL;
+   gl_shader_stage stage = info ? info->stage : MESA_SHADER_COMPUTE;
+
+   if (shader && shader->is_gs_copy_shader)
       return shader->selector->screen->ge_wave_size;
 
-   if (shader->selector->info.stage <= MESA_SHADER_GEOMETRY) {
-      return si_get_wave_size(shader->selector->screen, shader->selector->info.stage,
-                              shader->key.ge.as_ngg,
-                              shader->key.ge.as_es);
+   if (stage <= MESA_SHADER_GEOMETRY) {
+      return si_get_wave_size(sscreen, stage, shader->key.ge.as_ngg, shader->key.ge.as_es);
    }
 
-   return si_get_wave_size(shader->selector->screen, shader->selector->info.stage,
-                           false, false);
+   return si_get_wave_size(sscreen, stage, false, false);
 }
 
 static inline void si_select_draw_vbo(struct si_context *sctx)
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
index d059f860629..73c3f040698 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
@@ -425,7 +425,7 @@ struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
 
    shader->selector = gs_selector;
    shader->is_gs_copy_shader = true;
-   shader->wave_size = si_get_shader_wave_size(shader);
+   shader->wave_size = si_get_shader_wave_size(sscreen, shader);
 
    si_llvm_context_init(&ctx, sscreen, compiler, shader->wave_size);
    ctx.shader = shader;
diff --git a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
index a6827092fe4..68466cd88d6 100644
--- a/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
+++ b/src/gallium/drivers/radeonsi/si_shaderlib_tgsi.c
@@ -134,7 +134,9 @@ void *si_create_dma_compute_shader(struct pipe_context *ctx, unsigned num_dwords
    if (!ureg)
       return NULL;
 
-   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH, sscreen->compute_wave_size);
+   unsigned default_wave_size = si_get_shader_wave_size(sscreen, NULL);
+
+   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH, default_wave_size);
    ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT, 1);
    ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH, 1);
 
@@ -160,7 +162,7 @@ void *si_create_dma_compute_shader(struct pipe_context *ctx, unsigned num_dwords
    /* If there are multiple stores, the first store writes into 0*wavesize+tid,
     * the 2nd store writes into 1*wavesize+tid, the 3rd store writes into 2*wavesize+tid, etc.
     */
-   ureg_UMAD(ureg, store_addr, blk, ureg_imm1u(ureg, sscreen->compute_wave_size * num_mem_ops),
+   ureg_UMAD(ureg, store_addr, blk, ureg_imm1u(ureg, default_wave_size * num_mem_ops),
              tid);
    /* Convert from a "store size unit" into bytes. */
    ureg_UMUL(ureg, store_addr, ureg_src(store_addr), ureg_imm1u(ureg, 4 * inst_dwords[0]));
@@ -175,7 +177,7 @@ void *si_create_dma_compute_shader(struct pipe_context *ctx, unsigned num_dwords
       if (is_copy && i < num_mem_ops) {
          if (i) {
             ureg_UADD(ureg, load_addr, ureg_src(load_addr),
-                      ureg_imm1u(ureg, 4 * inst_dwords[i] * sscreen->compute_wave_size));
+                      ureg_imm1u(ureg, 4 * inst_dwords[i] * default_wave_size));
          }
 
          values[i] = ureg_src(ureg_DECL_temporary(ureg));
@@ -189,7 +191,7 @@ void *si_create_dma_compute_shader(struct pipe_context *ctx, unsigned num_dwords
       if (d >= 0) {
          if (d) {
             ureg_UADD(ureg, store_addr, ureg_src(store_addr),
-                      ureg_imm1u(ureg, 4 * inst_dwords[d] * sscreen->compute_wave_size));
+                      ureg_imm1u(ureg, 4 * inst_dwords[d] * default_wave_size));
          }
 
          struct ureg_dst dst = ureg_writemask(dstbuf, u_bit_consecutive(0, inst_dwords[d]));
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.cpp b/src/gallium/drivers/radeonsi/si_state_shaders.cpp
index 478e8d5bc85..07bb806f857 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.cpp
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.cpp
@@ -2270,7 +2270,7 @@ static bool si_check_missing_main_part(struct si_screen *sscreen, struct si_shad
          main_part->key.ge.as_ngg = key->ge.as_ngg;
       }
       main_part->is_monolithic = false;
-      main_part->wave_size = si_get_shader_wave_size(main_part);
+      main_part->wave_size = si_get_shader_wave_size(sscreen, main_part);
 
       if (!si_compile_shader(sscreen, compiler_state->compiler, main_part,
                              &compiler_state->debug)) {
@@ -2447,7 +2447,7 @@ current_not_ready:
 
    shader->selector = sel;
    *((SHADER_KEY_TYPE*)&shader->key) = *key;
-   shader->wave_size = si_get_shader_wave_size(shader);
+   shader->wave_size = si_get_shader_wave_size(sscreen, shader);
    shader->compiler_ctx_state.compiler = &sctx->compiler;
    shader->compiler_ctx_state.debug = sctx->debug;
    shader->compiler_ctx_state.is_debug_context = sctx->is_debug;
@@ -2714,7 +2714,7 @@ static void si_init_shader_selector_async(void *job, void *gdata, int thread_ind
            sel->info.stage == MESA_SHADER_TESS_EVAL || sel->info.stage == MESA_SHADER_GEOMETRY))
          shader->key.ge.as_ngg = 1;
 
-      shader->wave_size = si_get_shader_wave_size(shader);
+      shader->wave_size = si_get_shader_wave_size(sscreen, shader);
 
       if (sel->nir) {
          if (sel->info.stage <= MESA_SHADER_GEOMETRY) {