We calculated offsets as pairs, but addressed them in the shader as single float values, while reading them as ivec2s. Also removes unused code (was provisionally added if cooperative matrices could be used, but that turned out to be impossible).
Unified · Split
| @@ -94,7 +94,7 @@ static void insert_horizontal_pass(FFVkSPIRVShader *shd, int nb_rows, int first, · 93 unmodified lines | ||
| 94 | 94 | GLSLC(2, #pragma unroll(1) ); |
| 95 | 95 | GLSLF(2, for (r = 0; r < %i; r++) { ,nb_rows); |
| 96 | 96 | GLSLC(3, prefix_sum = DTYPE(0); ); |
| 97 | − GLSLC(3, offset = uint64_t(int_stride)*(pos.y + r)*T_ALIGN; ); | |
| 97 | + GLSLC(3, offset = int_stride * uint64_t(pos.y + r); ); | |
| 98 | 98 | GLSLC(3, dst = DataBuffer(uint64_t(integral_data) + offset); ); |
| 99 | 99 | GLSLC(0, ); |
| 100 | 100 | GLSLF(3, for (pos.x = 0; pos.x < width[%i]; pos.x++) { ,plane); |
| @@ -122,7 +122,7 @@ static void insert_vertical_pass(FFVkSPIRVShader *shd, int nb_rows, int first, i · 21 unmodified lines | ||
| 122 | 122 | GLSLC(0, ); |
| 123 | 123 | GLSLF(1, if (pos.x < width[%i]) { ,plane); |
| 124 | 124 | GLSLF(2, for (pos.y = 0; pos.y < height[%i]; pos.y++) { ,plane); |
| 125 | − GLSLC(3, offset = uint64_t(int_stride)*pos.y*T_ALIGN; ); | |
| 125 | + GLSLC(3, offset = int_stride * uint64_t(pos.y); ); | |
| 126 | 126 | GLSLC(3, dst = DataBuffer(uint64_t(integral_data) + offset); ); |
| 127 | 127 | GLSLC(0, ); |
| 128 | 128 | GLSLC(3, #pragma unroll(1) ); |
| @@ -167,40 +167,26 @@ static void insert_weights_pass(FFVkSPIRVShader *shd, int nb_rows, int vert, · 38 unmodified lines | ||
| 167 | 167 | GLSLC(0, ); |
| 168 | 168 | GLSLC(3, lt = ((pos.x - p) < 0) || ((pos.y - p) < 0); ); |
| 169 | 169 | GLSLC(0, ); |
| 170 | − if (TYPE_ELEMS == 4) { | |
| 171 | − GLSLF(3, src[0] = texture(input_img[%i], pos + offs[0])[%i]; ,plane, comp); | |
| 172 | − GLSLF(3, src[1] = texture(input_img[%i], pos + offs[1])[%i]; ,plane, comp); | |
| 173 | − GLSLF(3, src[2] = texture(input_img[%i], pos + offs[2])[%i]; ,plane, comp); | |
| 174 | − GLSLF(3, src[3] = texture(input_img[%i], pos + offs[3])[%i]; ,plane, comp); | |
| 175 | − } else { | |
| 176 | − for (int i = 0; i < 16; i++) | |
| 177 | − GLSLF(3, src[%i][%i] = texture(input_img[%i], pos + offs[%i])[%i]; | |
| 178 | − ,i / 4, i % 4, plane, i, comp); | |
| 179 | − | |
| 180 | − } | |
| 170 | + GLSLF(3, src[0] = texture(input_img[%i], pos + offs[0])[%i]; ,plane, comp); | |
| 171 | + GLSLF(3, src[1] = texture(input_img[%i], pos + offs[1])[%i]; ,plane, comp); | |
| 172 | + GLSLF(3, src[2] = texture(input_img[%i], pos + offs[2])[%i]; ,plane, comp); | |
| 173 | + GLSLF(3, src[3] = texture(input_img[%i], pos + offs[3])[%i]; ,plane, comp); | |
| 181 | 174 | GLSLC(0, ); |
| 182 | 175 | GLSLC(3, if (lt == false) { ); |
| 183 | − GLSLC(4, a = integral_data.v[(pos.y - p)*int_stride + pos.x - p]; ); | |
| 184 | − GLSLC(4, c = integral_data.v[(pos.y - p)*int_stride + pos.x + p]; ); | |
| 185 | − GLSLC(4, b = integral_data.v[(pos.y + p)*int_stride + pos.x - p]; ); | |
| 186 | − GLSLC(4, d = integral_data.v[(pos.y + p)*int_stride + pos.x + p]; ); | |
| 176 | + GLSLC(3, offset = int_stride * uint64_t(pos.y - p); ); | |
| 177 | + GLSLC(3, dst = DataBuffer(uint64_t(integral_data) + offset); ); | |
| 178 | + GLSLC(4, a = dst.v[pos.x - p]; ); | |
| 179 | + GLSLC(4, c = dst.v[pos.x + p]; ); | |
| 180 | + GLSLC(3, offset = int_stride * uint64_t(pos.y + p); ); | |
| 181 | + GLSLC(3, dst = DataBuffer(uint64_t(integral_data) + offset); ); | |
| 182 | + GLSLC(4, b = dst.v[pos.x - p]; ); | |
| 183 | + GLSLC(4, d = dst.v[pos.x + p]; ); | |
| 187 | 184 | GLSLC(3, } ); |
| 188 | 185 | GLSLC(0, ); |
| 189 | 186 | GLSLC(3, patch_diff = d + a - b - c; ); |
| 190 | − if (TYPE_ELEMS == 4) { | |
| 191 | − GLSLF(3, w = exp(patch_diff * strength[%i]); ,dst_comp); | |
| 192 | − GLSLC(3, w_sum = w[0] + w[1] + w[2] + w[3]; ); | |
| 193 | − GLSLC(3, sum = dot(w, src*255); ); | |
| 194 | − } else { | |
| 195 | − for (int i = 0; i < 4; i++) | |
| 196 | − GLSLF(3, w[%i] = exp(patch_diff[%i] * strength[%i]); ,i,i,dst_comp); | |
| 197 | − for (int i = 0; i < 4; i++) | |
| 198 | − GLSLF(3, w_sum %s w[%i][0] + w[%i][1] + w[%i][2] + w[%i][3]; | |
| 199 | − ,!i ? "=" : "+=", i, i, i, i); | |
| 200 | − for (int i = 0; i < 4; i++) | |
| 201 | − GLSLF(3, sum %s dot(w[%i], src[%i]*255); | |
| 202 | − ,!i ? "=" : "+=", i, i); | |
| 203 | − } | |
| 187 | + GLSLF(3, w = exp(patch_diff * strength[%i]); ,dst_comp); | |
| 188 | + GLSLC(3, w_sum = w[0] + w[1] + w[2] + w[3]; ); | |
| 189 | + GLSLC(3, sum = dot(w, src*255); ); | |
| 204 | 190 | GLSLC(0, ); |
| 205 | 191 | if (t > 1) { |
| 206 | 192 | GLSLF(3, atomicAdd(weights_%i[pos.y*ws_stride[%i] + pos.x], w_sum); ,dst_comp, dst_comp); |
| @@ -220,8 +206,8 @@ typedef struct HorizontalPushData { · 13 unmodified lines | ||
| 220 | 206 | int32_t patch_size[4]; |
| 221 | 207 | float strength[4]; |
| 222 | 208 | VkDeviceAddress integral_base; |
| 223 | − uint32_t integral_size; | |
| 224 | − uint32_t int_stride; | |
| 209 | + uint64_t integral_size; | |
| 210 | + uint64_t int_stride; | |
| 225 | 211 | uint32_t xyoffs_start; |
| 226 | 212 | } HorizontalPushData; |
| 227 | 213 | |
| @@ -275,8 +261,8 @@ static av_cold int init_weights_pipeline(FFVulkanContext *vkctx, FFVkExecPool *e · 47 unmodified lines | ||
| 275 | 261 | GLSLC(1, ivec4 patch_size; ); |
| 276 | 262 | GLSLC(1, vec4 strength; ); |
| 277 | 263 | GLSLC(1, DataBuffer integral_base; ); |
| 278 | − GLSLC(1, uint integral_size; ); | |
| 279 | − GLSLC(1, uint int_stride; ); | |
| 264 | + GLSLC(1, uint64_t integral_size; ); | |
| 265 | + GLSLC(1, uint64_t int_stride; ); | |
| 280 | 266 | GLSLC(1, uint xyoffs_start; ); |
| 281 | 267 | GLSLC(0, }; ); |
| 282 | 268 | GLSLC(0, ); |
| @@ -371,13 +357,11 @@ static av_cold int init_weights_pipeline(FFVulkanContext *vkctx, FFVkExecPool *e · 88 unmodified lines | ||
| 371 | 357 | GLSLF(1, ivec2 offs[%i]; ,TYPE_ELEMS); |
| 372 | 358 | GLSLC(0, ); |
| 373 | 359 | GLSLC(1, int invoc_idx = int(gl_WorkGroupID.z); ); |
| 374 | − | |
| 375 | − GLSLC(1, offset = uint64_t(integral_size)*invoc_idx; ); | |
| 376 | − GLSLC(1, dst = DataBuffer(uint64_t(integral_data) + offset); ); | |
| 377 | − | |
| 360 | + GLSLC(0, ); | |
| 361 | + GLSLC(1, offset = integral_size * invoc_idx; ); | |
| 378 | 362 | GLSLC(1, integral_data = DataBuffer(uint64_t(integral_base) + offset); ); |
| 379 | − for (int i = 0; i < TYPE_ELEMS*2; i += 2) | |
| 380 | − GLSLF(1, offs[%i] = xyoffsets[xyoffs_start + 2*%i*invoc_idx + %i]; ,i/2,TYPE_ELEMS,i); | |
| 363 | + for (int i = 0; i < TYPE_ELEMS; i++) | |
| 364 | + GLSLF(1, offs[%i] = xyoffsets[xyoffs_start + %i*invoc_idx + %i]; ,i,TYPE_ELEMS,i); | |
| 381 | 365 | GLSLC(0, ); |
| 382 | 366 | GLSLC(1, DTYPE a; ); |
| 383 | 367 | GLSLC(1, DTYPE b; ); |
| @@ -759,7 +743,7 @@ static int nlmeans_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) · 375 unmodified lines | ||
| 759 | 743 | /* Integral */ |
| 760 | 744 | AVBufferRef *integral_buf = NULL; |
| 761 | 745 | FFVkBuffer *integral_vk; |
| 762 | − uint32_t int_stride; | |
| 746 | + size_t int_stride; | |
| 763 | 747 | size_t int_size; |
| 764 | 748 | |
| 765 | 749 | /* Weights/sums */ |
| @@ -787,8 +771,8 @@ static int nlmeans_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) · 21 unmodified lines | ||
| 787 | 771 | return AVERROR(EINVAL); |
| 788 | 772 | |
| 789 | 773 | /* Integral image */ |
| 790 | − int_stride = s->pl_weights.wg_size[0]*s->pl_weights_rows; | |
| 791 | − int_size = int_stride * int_stride * TYPE_SIZE; | |
| 774 | + int_stride = s->pl_weights.wg_size[0]*s->pl_weights_rows*TYPE_SIZE; | |
| 775 | + int_size = s->pl_weights.wg_size[0]*s->pl_weights_rows*int_stride; | |
| 792 | 776 | |
| 793 | 777 | /* Plane dimensions */ |
| 794 | 778 | for (int i = 0; i < desc->nb_components; i++) { |
| @@ -982,9 +966,9 @@ static int nlmeans_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) · 187 unmodified lines | ||
| 982 | 966 | { s->patch[0], s->patch[1], s->patch[2], s->patch[3] }, |
| 983 | 967 | { s->strength[0], s->strength[1], s->strength[2], s->strength[2], }, |
| 984 | 968 | integral_vk->address, |
| 985 | − int_size, | |
| 986 | − int_stride, | |
| 987 | − offsets_dispatched * 2, | |
| 969 | + (uint64_t)int_size, | |
| 970 | + (uint64_t)int_stride, | |
| 971 | + offsets_dispatched, | |
| 988 | 972 | }; |
| 989 | 973 | |
| 990 | 974 | if (offsets_dispatched) { |